YC Fall 2026 (F26) companies

0.security builds an autonomous AI system that continuously finds and fixes security vulnerabilities for developers, AI-native startups, and enterprises, replacing manual penetration testing and periodic security audits. It surfaces exploit chains, prioritizes issues by business impact, and delivers validated remediation, covering work that traditional tools and human security teams handle on a slower, shallower cycle. Its research has produced fixes for critical CVEs in the Linux kernel, widely used open-source packages, and software running across government and commercial infrastructure.

Acyclic Labs builds infrastructure for running agent swarms that tackle software tasks too large for a single coding agent. Engineers point it at a repository and describe a task; the swarm splits the work across parallel agents, each running in its own sandbox, and merges the output into a patch or pull request. It replaces the manual effort of coordinating multiple AI coding sessions and stitching their results together.

Adside.ai builds a swarm of AI agents that runs paid ad campaigns across Meta, Google, and LinkedIn for startups and brands. The agents handle market research, strategy, ad creation, creative resizing, and ongoing account monitoring, replacing the work typically done by an in-house paid media team or an external agency. Clients can manage campaigns themselves through the platform or hand execution to a human expert.

Agent 37 provides persistent, always-on cloud sandboxes for AI agents, built for developers shipping agent-powered products. Each API call provisions an isolated sandbox per end user, pre-loaded with agents such as Claude Code, Hermes, and Codex, with files, memory, and connected tool credentials that persist across sessions. This replaces the need to configure servers, manage cloud infrastructure, or rebuild agent state between runs.

Agent Relay builds coordination infrastructure for teams running multiple AI coding agents in parallel. It gives agents such as Claude Code, Codex, and OpenCode a shared messaging layer, persistent session history, and direct connections to tools like GitHub, Linear, and Slack, replacing the manual work of tracking, handoff, and context-switching across separate agent sessions. Engineering teams get a single plane to monitor every agent, enforce human approval steps, and run resumable multi-agent workflows on cloud, local, or self-hosted infrastructure.

Ageospatial builds a geospatial AI platform that reads satellite imagery, street-level imagery, maps, real-time data, and public records to profile physical assets, including commercial properties, land, and infrastructure. Insurers, carriers, and MGAs use it to assess what a property actually is, what hazards surround it, and how it has changed over time, replacing the manual physical inspection that most properties never receive. Each property becomes a continuously updated record that underwriting, claims, and pricing teams can audit from past to present.

Aikido Technologies builds offshore floating datacenters that co-locate AI compute with on-platform wind generation and ocean cooling. Each platform integrates a 15MW+ wind turbine, battery storage, and 10MW+ of prefabricated data halls inside a modular steel structure assembled quayside and moored at sea. The product targets hyperscale and sovereign datacenter operators who need gigawatt-scale capacity fast, replacing years of onshore land acquisition, grid connection, and cooling infrastructure work.

Altair Health runs end-to-end medical billing for independent physician practices, replacing in-house billing staff and outsourced billing services. An AI agent handles claim submission, prior authorizations, eligibility checks, and denial appeals by checking every action against a formal, prover-verified model of each insurer's current policies. Human billers step in only when the prover flags an exception, so claims go out clean and denials are worked immediately rather than queued.

antimattr builds a voice-first hardware system, comprising a ring and earphones, for people who want to capture thoughts, dictate text, and control AI agents hands-free. The earphones apply proprietary voice isolation to strip background noise so only the speaker's voice comes through, while the ring lets users record conversations, log ideas on the go, and remotely trigger agents such as Claude Code and Codex. The product replaces manual note-taking, keyboard dictation, and direct on-device agent interaction.

Antropi Robotics builds autonomous CNC factory systems that replace manual machine operation, material handling, and maintenance scheduling. Gantry-mounted robotic arms with built-in vision models load and unload CNC machines, autonomous vehicles move raw materials and finished parts between stations, and IoT sensors feed an AI layer that predicts failures and schedules preventive maintenance before downtime occurs. The target customer is a discrete parts manufacturer running CNC equipment who currently relies on human operators for machine tending, inter-machine logistics, and reactive maintenance.

AnyAlpha builds an agentic trading platform that converts a user's trading thesis into an autonomous agent, starting with prediction markets. The agent researches the market, backtests the strategy against historical data, and executes live trades around the clock across multiple exchanges. It replaces the manual work of strategy research, backtesting, and trade execution that retail traders otherwise do themselves.

Applied Kinetics builds AI agents for energy and data center construction teams, handling the coordination work that keeps long-lead equipment orders on track. The agents chase suppliers, reconcile engineering changes, and monitor delivery milestones by connecting directly to ERP systems, email, spreadsheets, and document stores, replacing the manual follow-up that project teams currently do across inboxes and spreadsheets. Results and exceptions surface in Microsoft Teams or Slack, where the team already works.

Archangel Health builds post-training and evaluation infrastructure for medical AI, targeting frontier labs and medical AI developers. It provides reinforcement learning environments inside de-identified EHRs, physician-verified training data, and clinical reasoning benchmarks designed around hard diagnostic cases. The product replaces ad-hoc fine-tuning pipelines and internal benchmark construction that labs would otherwise build themselves.

Archgen Labs builds AI-powered electronic design automation tools for 3D semiconductor physical design. Its core product, Newton, is an autonomous agent that handles placement, routing, design-rule fixes, and timing closure, replacing the iterative manual triage that chip engineers do today. A companion thermal analysis tool, LayerTherm, predicts on-chip temperatures orders of magnitude faster than conventional simulation, letting teams catch hotspots and compare more design options earlier in the flow.

Armorer Labs builds a security platform for companies in regulated industries that deploy AI agents. It maps what an agent can be manipulated into doing, simulates attacks to reproduce real vulnerabilities, and writes verified fixes ready for engineer approval, typically within hours. In production, it inspects every agent action in milliseconds, blocking prompt injection, goal hijacking, and data exfiltration without slowing the agent for end users.

Asakana automates order entry for food distributors, converting orders that arrive by text, email, WhatsApp, voicemail, and WeChat directly into drafted ERP records. Restaurants routinely send orders through these informal channels, and distributor staff would otherwise retype each one manually. Anything ambiguous is flagged for human review before it posts, keeping the data clean without adding a bottleneck.

Ascenix builds a clinical intelligence platform for skilled nursing and long-term care facilities. It reads the unstructured notes nurses already enter in EHRs like PointClickCare and converts that data into automated incident reporting, compliance gap alerts, reimbursement recovery, and forward-looking resident risk rankings. The product replaces manual chart reviews and spreadsheet-based reporting that clinical and operational teams otherwise build by hand.

Asim builds safety research and testing infrastructure for robotics teams. It runs robots through real-world and simulated scenarios to measure physical harm, semantic misunderstanding, and malicious-use risks, benchmarking results against a human baseline. The work replaces ad-hoc internal safety validation with structured audits and a runtime layer that monitors, predicts, and blocks unsafe actions during execution.

Aucert deploys a bench of specialized AI agents that verify mobile apps before each release, replacing the weeks of manual QA work that top apps currently require from multiple human teams. Each agent covers a distinct domain, including functional quality, accessibility, and app store policy compliance, and completes a full review in hours. Aucert plugs into agentic coding workflows so detected issues can be resolved within the same development loop.

Audora converts written stories into multi-voice audiobooks, assigning a distinct, personality-matched voice to each character and a separate voice to the narrator. It targets readers of fanfiction, self-published fiction, and web serials on platforms like AO3, Wattpad, and Substack, producing audio for stories that professional studios never record. The product replaces flat, single-narrator text-to-speech tools and the manual work of casting and producing a full-cast audiobook.

Aviern builds fixed-wing unmanned aircraft that autonomously patrol infrastructure corridors and return geospatial intelligence to operators. The aircraft launch and land from unattended docks installed along the right-of-way, with onboard perception that detects pipeline leaks, encroachments, vegetation stress, and ground movement, replacing the helicopter crews operators currently hire for mandated patrol flights. The same aircraft and dock platform extends from pipelines to power grids, rail, borders, and military bases.

Axonnic builds a wearable EMG sleeve that captures muscle signals, finger positions, and applied force as workers perform dexterous tasks by hand. Manufacturers use these recordings to train robots on physical jobs that video-based data collection cannot document accurately, replacing manual demonstration capture and conventional motion-tracking setups. The same signal pipeline also enables human-computer interaction driven directly by muscle activity rather than cameras or controllers.

Bayworks builds AI-powered software for automotive service bays, helping technicians diagnose and repair vehicles faster at a time when US shops face a severe shortage of qualified technicians. Each repair captured feeds an intelligence layer that automates broader service operations, replacing manual coordination and tribal knowledge with a continuously learning system. The product works across all major vehicle makes.

Blinka builds a USD banking platform for US-incorporated companies whose operations, teams, and vendors are based outside the United States. It replaces a patchwork of traditional bank accounts and wire services by offering USD accounts, global corporate cards with no foreign-transaction fees, cross-border payments without standard FX markups, and treasury yield on idle balances, all in one platform. The AI-powered back-office tools handle the financial operations work that would otherwise fall to manual processes or separate software.

Boat provides cloud-hosted Ubuntu virtual machine sandboxes built for AI agents. Each sandbox ships pre-configured with Docker, Chrome, common languages, a virtual desktop, and snapshot-based forking, so agent developers avoid standing up and maintaining their own compute infrastructure. The product targets teams running agents at scale on code, research, legal, and other knowledge work.

Boxes.dev gives AI coding agents, specifically Claude Code and Codex, dedicated cloud Linux machines so they can run in parallel without competing for local resources. Developers trigger agents from Slack, Linear, a schedule, or an API call, then monitor or steer them from a desktop app, mobile device, or terminal. It replaces local worktree setups and the need to keep a laptop open while agents work.

ByteAsk builds an AI coding agent purpose-built for C and C++, the languages at the core of trading systems, automotive software, chip firmware, and operating systems. Rather than generating a plausible-looking diff and stopping, it runs the actual compiler, sanitizers, debugger, and test suite against every change until the output is verified clean. Engineers replace the manual cycle of writing code, building, running tests, and triaging failures by hand.

Canos builds a liquid biopsy platform that detects cancer in dogs from a single blood draw, sequencing cell-free DNA to analyze fragmentation, copy number variation, and methylation signals. Healthy dogs get a genomic baseline banked at the time of testing, updated annually to track any emerging signals. Dogs already diagnosed with cancer use a separate monitoring product to follow molecular changes through treatment.

Capveon builds an AI layer for infrastructure contractors, manufacturers, and supply chains, sitting on top of existing systems like Salesforce, Viewpoint, HeavyBid, and Dodge without replacing them. It automates the work that falls between those systems: bid qualification, preconstruction research, spec review, quote pricing, and forecasting. The manual effort it replaces ranges from retyping data across CRM and ERP tools to reading specifications and assembling owner and job histories before preconstruction kickoffs.

Cardinal Square runs an AI-native supplemental school, pairing AI with real educators to deliver personalized daily practice in reading, writing, and math for K-12 students. It targets the same ground as tutoring chains like Kumon and Mathnasium, replacing their in-person worksheet-and-tutor model with 30-minute AI-guided sessions online. The longer-term plan is to expand from supplemental coursework into full accredited schools.

Caveman builds an optimization stack for AI agents that cuts inference and runtime costs without degrading agent quality. It monitors live agent workloads, identifies cost and performance inefficiencies, tests fixes, and rolls out only the improvements that pass validation. The product replaces manual prompt tuning, ad-hoc model-swapping, and guesswork that engineering teams currently do to control agent spend.

Cesara builds AI agents that help health system billing teams find and recover underpayments from insurers. The platform ingests payer contracts, claims, and remittance data to answer contract and payment questions in seconds, replacing work that would otherwise take billing staff hours or months. When a claim is paid below the contracted rate, Cesara drafts the reconsideration request, citing the specific contract clause and dollar discrepancy.

ChemPilot.ai accelerates the discovery of novel materials for defense and dual-use applications, including energetics, propellants, carbon fibers, specialty polymers, coatings, and composites. Its customers are defense primes and government R&D agencies that today spend over a decade on trial-and-error lab work to qualify a new material. ChemPilot.ai replaces that process with physics-based simulation, machine learning, and AI agents that screen candidates for synthesizability before anything reaches a lab, compressing development timelines from years to months.

ClaimKit prepares federal and state R&D tax credit claims for hardware tech companies, combining AI-generated documentation with review by former R&D tax consultants. It targets the work traditionally done by accounting firms and specialist consultants: gathering evidence, identifying qualifying research expenses, and producing the filings. Finance teams and founders use it instead of engaging an external firm for the same engagement.

Clean surfaces buying signals for B2B sales teams by monitoring real commercial activity, such as customs shipments, permits, loans, and supplier changes, rather than ad clicks or web traffic. It tells sellers which companies are actively in a buying cycle right now, replacing manual prospect research and gut-feel timing. The target customer is any B2B revenue team that needs to prioritize outreach by actual market behavior.

Collar builds an AI-native investor relations platform for undercovered public companies. It automates earnings preparation, compiles peer comparisons and question banks, and ranks investors by portfolio fit so CFOs and IR teams spend less time on research and more time in front of aligned institutions. The product replaces weeks of manual work: drafting earnings packets, scouring 13F filings, and building roadshow target lists by hand.

Composal builds automated pre-release testing for software teams shipping products with AI agents. It deploys a swarm of agents that navigate an application as real users would, exercising critical flows and edge cases in parallel. Teams get video recordings, logs, and reproduction steps for every failure, replacing manual click-through testing before each release.

Confetti lets anyone design and build physical hardware projects through plain-English conversation, replacing weeks of CAD learning, parts sourcing, and firmware writing. Users describe what they want, iterate in chat, and receive compatible components and 3D-printed parts shipped to their door. The product targets makers and non-engineers who want a finished physical object without prior technical skills.

CONUS builds an AI knowledge platform for operating rooms, serving surgeons, OR nurses, and medical device sales representatives. It centralizes product information from device manufacturers, hospital protocols, and medical research into a single searchable layer, replacing scattered documentation and manual lookup during surgical workflows. The platform also gives device manufacturers a direct channel to place products with hospital end users.

Covera builds AI agents that handle the operational workload inside insurance brokerages: collecting submission documents, entering risk profiles into carrier portals, comparing returned quotes, triaging claims, chasing unpaid premiums, and identifying coverage gaps across a book. Each agent takes on one discrete task and delivers a ready-to-approve output, replacing the manual, repetitive work that keeps brokers away from clients.

Cyclon builds a smartphone where AI serves as the primary interface, replacing the conventional app-and-screen model. Instead of navigating menus and tapping through applications, users state what they want and the phone dispatches AI agents to control apps or run tasks in the background on their behalf. The product targets everyday smartphone users and replaces the manual, multi-step interaction that current mobile operating systems require.

Databuddy builds cookieless product analytics for startups and developer teams that lack a dedicated data analyst. It tracks visitors, funnels, signups, errors, and web vitals in one place, and surfaces a built-in AI analyst called Databunny that watches metrics continuously, flags meaningful changes, and explains the cause behind each shift. This replaces the manual work of checking dashboards and diagnosing product issues such as conversion drops, user frustration, and churn.

Decent builds a health insurance brokerage for small businesses that replaces the once-a-year broker review with continuous plan management. It quotes every available coverage option across carriers and PEOs, then monitors claims data throughout the year to negotiate renewals and surface cheaper or richer plans before the employer ever asks. The work it replaces is a passive broker who shops coverage once at renewal and does nothing in between.

DeepMark embeds a machine-readable identity signal directly into AI-generated audio at the moment of production, surviving VoIP, codecs, and the phone network intact. Banks, call centers, and end recipients can verify in real time that a call originated from an authorized AI agent, replacing trust that previously depended on metadata that gets stripped in transit. The same embed-and-verify flow also marks AI-generated images and video for regulatory compliance, including EU AI Act transparency obligations.

Definite builds a shared data layer for finance teams, connecting ERPs, ledgers, payroll, and payment systems into a single, continuously updated model of a company's books. On top of that model, back-office teams build and run automation agents using plain English, without writing code or pulling in engineers to stitch data sources together. The product replaces the manual, per-workflow data plumbing that currently blocks finance teams from automating repetitive back-office work.

Degla Inc builds an autonomous drone operations system for public safety and security teams. Operators type a plain-language goal into SkyCommander, and Degla's onboard intelligence plans the mission, launches a purpose-built 4.4-meter fixed-wing VTOL, and adapts in flight, replacing manual flight planning, continuous operator control, and separate sensor management. The work it displaces is hands-on drone piloting and real-time replanning by human crews.

Deploy puts generalist robots on warehouse stations handling returns processing, kitting, autobagging, and FBA prep, replacing human labor at the tasks logistics operators struggle to staff. The company handles everything between the robot's underlying model and the physical job, delivering guaranteed throughput from day one across industries that share the same deployment challenge.

Dreamscale Labs builds a cloud inference platform for physical AI models used in robotics. Robotics teams use it to offload model inference from on-robot GPUs to the cloud, running models like NVIDIA's DreamZero and MolmoAct2 with real-time latency. The platform replaces costly dedicated on-device compute by optimizing the model, GPU kernels, and network stack together.

Dydema builds AI models that decode brain signals and collects brain recordings on demand, so any team developing brain-computer interfaces can train and run their own neurotech software. The models handle signal interpretation that developers would otherwise have to build from scratch, and the data service fills a gap that currently limits the whole field. Together, they replace the need for neurotech teams to run their own data collection programs or build low-level signal-processing pipelines.

Elimentary builds AI agents that run middle and back-office operations for private credit managers, handling tasks such as loan tape analysis, borrowing base calculations, covenant monitoring, waterfall automation, and NAV calculations. The product replaces the spreadsheets, emails, and manual analyst hours that most private capital firms still rely on for post-investment operations. It ingests loan tapes, deal documents, and servicer reports and produces decision-ready data, end to end, without human intervention.

Epsilab builds reinforcement learning environments derived from real-world data, used to train AI agents on long-horizon, multi-step tasks. It targets applied AI research teams and frontier model developers who need realistic simulation grounds that go beyond synthetic or narrow benchmarks. The product replaces hand-crafted, toy RL environments with simulations grounded in actual workflows and multi-agent dynamics.

ernd builds a corporate credit card for startups that spend heavily on AI infrastructure, including inference and GPU costs. It pools cardholders' collective spend to negotiate better pricing, higher usage limits, and priority capacity from AI providers, replacing the individual negotiations and standard rates a startup would face on its own. Cardholders also get vendor support escalation and travel booking handled by the ernd team.

OpenSEO builds an open-source SEO platform that gives founders and marketers keyword research, backlink analysis, rank tracking, site auditing, and AI-brand visibility tools, replacing paid subscriptions to tools like Semrush and Ahrefs. It also exposes SEO data directly to AI agents via an MCP server, letting tools such as Claude and Codex run keyword research, audits, and competitor analysis without a human in the loop.

Buildbox benchmarks AI models on physical science problems, working with scientists to test how models reason, use specialized tools, and interpret experimental evidence. It targets researchers who need to judge where AI can contribute reliably and where its results require further scrutiny, replacing ad hoc, informal assessments of AI capability with structured, reproducible evaluations grounded in real research problems.

Faraday builds an operations platform for manufacturers and distributors that coordinates back office and shop floor work across AI agents, robots, and human workers in a single runtime. It connects LLM models, an agent harness, operational data, business applications, and physical robots so that planning, execution, and record updates stay synchronized end to end. The platform replaces fragmented point tools and manual coordination across inventory management, production scheduling, and related industrial workflows.

Fire Wallet builds a non-custodial, end-to-end encrypted stablecoin wallet with native payment processing built in. Users hold their own keys, route payments at lowest cost without a third-party processor, and move between stablecoins and fiat directly inside the app. The product targets individuals and businesses that currently hand custody and payment routing to centralized wallet and payments providers.

Fleety builds AI-powered dispatch software for plumbing and HVAC companies running high volumes of daily jobs. It listens to inbound calls, validates booking details against equipment records and past notes, then routes each job to the technician best positioned to generate revenue, replacing manual dispatcher judgment. Alongside dispatch, it handles after-hours call answering, lead qualification, and technician performance analytics.

Forward builds an AI origination platform for private credit, serving commercial debt advisors and private credit funds. Advisors submit raw borrower materials once and the platform parses financials, structures a deal proposal, and matches the packaged opportunity to funds whose mandates fit. It replaces the manual work of spreading financials, writing credit memos, and cold-sourcing lender relationships.

Frontrunner builds AI agents that handle go-to-market work for B2B sales and marketing teams: researching accounts, finding prospects, running campaigns, qualifying inbound leads, and keeping the CRM current. The agents connect to a team's existing tools and work alongside human reps, replacing the manual, repetitive execution that consumes GTM headcount.

GodHands reconstructs an outgoing employee's role, in full, directly from the systems they worked in, so enterprises can hand that role over without disruption. It maps what depends on the person, surfaces the knowledge embedded in their day-to-day work, and keeps operations running after they leave. The product replaces the manual, error-prone process of knowledge transfer that typically stretches across weeks of documentation and shadowing.

Vendo gives B2B software products a customization layer so end users can build their own features, micro-apps, and automations directly inside the product, without filing requests or waiting on an engineering backlog. It reads a company's existing repo, API, and component library, then renders everything it generates inside that product's own brand and security sandbox. The open-source layer ships with a single install command and replaces the manual work of triaging, scoping, and building one-off feature requests for individual customers.

Graphide builds a code editor for software teams whose AI agents are writing code faster than developers can review it. Instead of presenting line-by-line diffs, it renders the entire codebase as a live graph of modules, classes, and functions, updating in real time as agents work. It checks each agent's changes against the intended architecture, replacing the manual work of tracing what actually changed across a system.

GrayPass verifies that the authenticated user is still in control of a session at the moment a consequential action executes, closing the gap between login and action. It analyzes cognitive biometrics, including keystroke timing, pointer movement, and scroll patterns, alongside session and device context to assess continuous control. Before an action such as a payment approval, permission change, sensitive data export, or AI agent execution proceeds, GrayPass returns an allow, challenge, hold, or deny decision and generates signed proof tied to that specific action.

Grep AI builds and deploys managed AI agent workforces for research-intensive knowledge work in financial services and enterprise. Customers including global banks, fintechs, and payment companies use its agents to run due diligence, compliance checks, credit research, sanctions screening, document reviews, and reporting. The agents replace manual analyst work that carries high accuracy requirements and significant business risk if done wrong.

HABU operates as an AI-native real estate developer, acting as a fee and merchant developer for capital investors who want housing projects delivered on time and at forecasted returns. It replaces the traditional development firm by running the full project lifecycle, from land screening and entitlement through construction management and disposition, using agentic workflows, geospatial analysis, and development-specific automation. The result is more projects per staff member, faster schedules, and lower total project costs, without HABU holding property long term.

Halmos Labs builds automated research software for biotech and drug discovery teams. It designs experiments by simulating every possible outcome across multiple models, then surfaces every assumption behind each recommendation so scientists know exactly what the system knows and where it is uncertain. The work it replaces is manual experiment planning and the opaque model outputs that force researchers to run more experiments than necessary to reach a conclusion.

Helios Intelligence builds an at-home blood test that measures more than 1,000 biomarkers from a single drop of blood and converts the results into a continuously updated model of the user's biology. The product targets health-conscious consumers and surfaces risk earlier than conventional clinical testing by combining molecular data with wearables, labs, and health records. It replaces the periodic, narrow-panel bloodwork ordered reactively by physicians with a longitudinal, high-resolution picture of what is happening inside the body before symptoms appear.

Hemlock builds force-annotated demonstration data for robotics AI training. Its Bracelet, a wrist-worn EMG band, reads forearm muscle activity to estimate finger-level grip force during human demonstrations, capturing what video and joint-angle data alone cannot: how hard the hand is pressing. Robotics labs use this to train manipulation policies for tasks such as parts assembly, delicate object handling, and hospital supply workflows, replacing the guesswork that comes from training on force-blind recordings.

herdr builds an open-source agent runtime for developers running fleets of coding agents across multiple machines. A persistent background server owns each agent's terminal session, keeping work alive through disconnects, closed laptops, and dropped SSH connections. Developers get a unified attention queue showing every agent's status, working, blocked, or done, across all projects, and can reconnect from any device without losing progress.

Hesper AI builds an end-to-end claims resolution platform for property and casualty carriers, MGAs, and TPAs. Its AI agents handle every stage from first notice of loss through coverage verification, fraud investigation, settlement, and subrogation recovery, replacing the manual review work that today stretches a typical claim past forty days. Routine claims close automatically; suspicious ones get a full investigative workup with an audit-ready evidence trail, while adjusters retain final decision authority.

Hickory & Company builds an AI-powered compliance and Bank Secrecy Act platform for banks, fintechs, and credit unions. Agents handle the repetitive review, documentation, and reporting work inside a financial institution's compliance program, cutting multi-hour case work down to brief reviews while keeping every decision with the human team. All outputs are documented and cited so they are ready for regulatory examination.

HillStar builds procurement intelligence software for utilities and energy infrastructure companies. It automates the full RFx lifecycle, from translating engineering scopes into sourcing strategies to evaluating vendor bids and modeling award scenarios, replacing months of manual procurement work by specialized teams. The platform captures past evaluator decisions and supplier outcomes to sharpen future procurement across complex, high-value capital spending programs.

Hopper builds a specialized inference platform for voice AI, serving companies that run speech-to-text, text-to-speech, and speech language models in production. It fine-tunes those models on a customer's own call traffic and continuously improves them against live usage, replacing generic hosted model APIs. The result is lower latency and lower cost per token than general-purpose alternatives.

horizontal. builds AI agents that run manufacturing operations end-to-end, sitting on top of the ERP, MES, QMS, and CMMS systems factories already use. The agents handle the coordination work that today falls to factory teams: replanning production after a disruption, tracing quality deviations to their source, and releasing batches as soon as checks clear. Target customers are enterprise manufacturers in process, batch, and discrete industries in Europe and the US.

Hue builds a testing platform for AI agent teams. It captures real production runs and converts them into repeatable test cases that preserve the original user context, tool calls, and stateful environment, so teams can replay and score agent behavior under conditions that match actual usage. The product replaces ad hoc manual testing and synthetic benchmarks that fail to reflect the complexity of live deployments.

Hundred builds an AI visual inspection system for contract manufacturers producing consumer electronics. It scans incoming parts in their shipping trays, grades each one against cosmetic specifications, and flags sub-millimeter defects, replacing human inspectors who check parts under microscopes. The system requires only a single golden sample to calibrate, with no labeled defect data, so it keeps pace with the rapid part revisions that defeat traditional vision systems.

Ignition Benefits brokers employee health insurance for companies under 200 employees, replacing the traditional brokerage model with an AI-native platform backed by licensed advisors. It gives employers visibility into their own workforce health costs, fights annual rate increases at renewal, and handles plan enrollment and administration. The product targets small and mid-market employers who have historically accepted opaque pricing and year-over-year premium hikes with no analytical recourse.

Inflection Space Systems builds high-volume spacecraft manufacturing infrastructure for commercial and government customers. The company targets the production bottleneck that limits fleet scale: today's spacecraft are built one at a time in specialized clean rooms, a process that cannot support next-generation constellations numbering in the tens of thousands. Inflection replaces that artisanal model with industrial-scale factory throughput.

Invertix deploys custom AI agents inside energy and critical infrastructure companies, ingesting SCADA data, work orders, and operational records to monitor assets, investigate faults, and coordinate work across operations teams. The models are grounded in physics-informed neural networks and differentiable simulators trained on data no general-purpose model has seen. This replaces manual monitoring, ad-hoc fault investigation, and fragmented coordination between field operators and back-office teams.

isoquant runs open-source AI models on its own inference cloud, optimized across the full stack from GPU kernels to serving orchestration. Engineering teams replace their existing inference providers by swapping in an isoquant API endpoint, getting lower latency and higher throughput at reduced per-token cost. The product targets companies running inference at scale who are paying too much or hitting speed limits with current providers.

Jinko builds travel infrastructure for developers shipping AI agents. It provides a single API covering flights, hotels, car rentals, and activities, letting agents handle the full trip lifecycle: open-ended search, live fare retrieval, booking, and post-booking changes. It replaces the need to stitch together multiple travel data providers and booking flows inside an agent.

Kailash Labs builds small, task-specific video understanding models for industries that process high-volume camera feeds, including security, industrial monitoring, robotics, and live broadcast. Its automated pipeline takes a customer's objective, distills a frontier vision-language model, and produces a fine-tuned model optimized for that single task, deployable on-premises where data privacy rules out cloud inference. The result replaces both expensive generic frontier models and brittle, single-purpose computer vision classifiers that cannot transfer across tasks.

Kairos Health builds an AI receptionist for dental practices that handles both inbound and outbound patient calls around the clock. It connects directly to a practice's existing management software, books appointments, follows up with patients overdue for care, and verifies insurance, replacing the work a front-desk coordinator or call-handling staff would otherwise do manually.

Kairox Systems builds autonomous hybrid drones that transition between fixed-wing flight and subsurface operation in a single mission. The platform, called Tempest, targets U.S. and allied militaries for littoral missions, including mine countermeasures, coastal surveillance, and one-way strike, that currently require separate air and underwater systems. A single vehicle can fly to an area of operations, submerge beneath a surface vessel, then breach the waterline and continue flight, replacing multi-vehicle mission packages with one low-cost, attritable airframe.

Keres builds reusable unmanned jet interceptors for air defense. Its vehicle, the Garmur, launches on command, closes on targets at jet speed, identifies and destroys drones and cruise missiles, then returns to land and fly again. The system gives military commanders a fighter-class intercept capability at a fraction of the cost of crewed or single-use alternatives.

Kiwi builds a data engineering layer for go-to-market AI agents, serving RevOps, BizOps, sales, and marketing teams. It syncs live data from tools like Salesforce, Gong, and HubSpot directly into each agent run, and cites every claim back to the exact SQL query and data snapshot that produced it. This replaces the manual work of keeping GTM agent context, connectors, and data pipelines accurate as strategies shift.

Lantern AI builds an end-to-end AI hiring manager for companies replacing the combined work of a recruiter, HR partner, and talent intelligence function. It takes a 15-minute calibration to learn what strong looks like at a specific team, pressure-tests that standard against the real talent market, then runs the full hiring loop: sourcing, screening, interviewing, assessing candidates, coordinating schedules, and extending offers. The product targets B2B customers who want a single, continuous system instead of handoffs between people and tools.

Lark builds AI agents for wholesale distributors, automating the operational work that purchasing teams, inventory managers, and sales staff handle manually today. The agents connect to a distributor's existing ERP, learn from historical transaction data and past team decisions, and handle purchasing, demand forecasting, inventory monitoring, sales, and analytics. A distributor can run its full back-office operation through Lark rather than through manual review and spreadsheet-based workflows.

Latent detects hallucinated LLM outputs in real time by reading a model's internal state as it generates each answer, before any response reaches an end user. Teams deploying LLMs in finance, legal, healthcare, and customer support get a prioritized review queue of flagged answers, each with a plain explanation, replacing the manual spot-checking or expensive judge-model pipelines those teams would otherwise run. The product sits inside the customer's own deployment, so no output leaves their environment to be verified.

Limnion builds open-standards telemetry storage for engineering teams running observability at scale, including agentic workflows. Its core project, Siglake, ingests OpenTelemetry data natively, stores it as Parquet and Iceberg on object storage, and queries it with DataFusion SQL, replacing proprietary telemetry backends and the vendor lock-in that comes with them. A managed bring-your-own-cloud platform and a streaming anomaly-detection layer are in development on top of that open core.

Liquid Energy builds self-contained, solar-powered AI inference modules called NanoDCs: factory-assembled data centers that ship whole, unfold their own solar arrays on arrival, and begin delivering inference tokens within eight weeks of contract signing. Each unit carries its own compute, power, cooling, and networking, removing the grid connections, construction timelines, and permitting processes that slow conventional data center builds. The product targets companies of any size that need to scale AI inference capacity faster than the grid allows.

Lux Foods engineers ultra-compact, fruit-dominant tomato plants that grow on sugar under minimal light, replacing conventional greenhouse and field farming. The plants are propagated through asexual cloning and cultivated in robotic vessel systems, removing sunlight, pesticides, and seasonal constraints from production. The result is year-round cherry tomato output sited close to urban consumers, with biology and robotic workflow designed together so the system scales by replicating a standardized unit.

Lyon trains foundation models from scratch on the proprietary event data of banks, insurance companies, and fintechs. Each model learns from a company's own transactions, payments, clicks, and customer interactions, then runs inside that company's cloud on AWS, GCP, or Azure. A single model powers every downstream prediction, from credit risk and fraud to churn, collections, income, and cross-sell, replacing separate feature-engineering projects for each task.

Mantle runs a full-stack field service operation that deploys, repairs, and maintains machines and robots on behalf of the hardware and robotics companies that build them. It takes direct ownership of uptime rather than acting as a marketplace or software layer, replacing the need for manufacturers to pull their own engineers onto customer support calls. The target customer is any company shipping physical machines that needs reliable post-sale service without building that capability in-house.

Maritime provides managed cloud infrastructure for AI agents, giving each end user their own isolated, persistent virtual machine. Developers building multi-agent products can provision thousands of per-user agents without writing any infrastructure code, replacing self-managed VM orchestration, scaling logic, and sleep/wake scheduling. Each agent retains its own filesystem, credentials, and browser state across sessions, so it can resume work after days of inactivity.

Melty lets PC gamers combine titles they already own into cross-game "mashups" and play them with friends in one click. It replaces the manual work of finding, installing, and configuring mods by handling setup automatically and generating invite links for multiplayer sessions. Creators build and publish their own mashups by feeding instructions to an AI coding agent such as Claude Code or Codex.

Micora builds palletized, containerized solar-plus-battery microgrids that ship anywhere in the continental United States within seven days and scale from 50 kW to multi-megawatt capacity. Forklift robots scan terrain and place solar pallets autonomously, assembling roughly one megawatt every eight days without specialized construction labor. The system targets data centers, advanced manufacturers, and other high-demand sites that need utility-scale power faster than conventional grid interconnection or EPC-built solar farms can deliver it.

Mirrors builds a runtime context layer for production AI agents. It monitors live agent sessions, learns from failures like context loss, infinite loops, and malformed tool calls, then automatically routes corrective context the next time the same situation arises. The product targets engineering teams operating agents that take irreversible actions, replacing manual trace review and hand-patched prompts after regressions.

Munito AI provides open-model inference for European enterprises through an API that routes requests to EU-based compute. It replaces reliance on US-hosted, closed-model providers by keeping both inference and data storage within the EU, with zero data retention and no training on customer data. The company was founded by the former inference lead at Mistral AI.

Native builds an autonomous marketing platform for small businesses that creates, designs, and publishes social media posts and paid ads across channels without requiring any marketing expertise from the owner. An in-house predictive model called Norn determines what content people will respond to. The platform replaces the work of a social media manager and ad buyer, handling copy, visuals, scheduling, and ad optimization end to end.

Neomatter builds an AI layer for in-house legal teams that intercepts requests from across a company before they reach counsel. It connects to email and messaging tools, reads incoming documents, checks the team's playbooks and past deals, and delivers a completed analysis with a recommended response for the lawyer to approve. The product replaces the intake, research, and first-draft work that in-house attorneys otherwise do themselves.

Nitroduck builds CRISPR-based gene editing tools for crop breeders, covering vegetables, fruits, and specialty crops that current platforms do not support. Breeders use Nitroduck to develop varieties with traits like heat tolerance, pollinator independence, and novel consumer appeal, replacing conventional breeding cycles and pesticide-dependent disease management.

Nodus Compute provides cloud infrastructure purpose-built for AI workloads, including model training, fine-tuning, and agent execution. AI engineers and teams use it to run pre-training, post-training, and LoRA fine-tuning jobs with their own code and data, replacing the need to manage general-purpose cloud compute. It also hosts sandboxed environments for AI agents, letting coding agents like Claude Code, Codex, and Cursor execute tools, run evaluations, and complete long-running tasks within configurable resource limits.

Nxtcure Labs runs clinical trials end to end through Galen, a platform of thirteen purpose-built AI agents that handle the routine operational work of a contract research organization, from site feasibility and activation through monitoring, safety reporting, and biostatistics. Every action with clinical, safety, regulatory, or financial consequence is held in a single human review queue and requires approval from a qualified person before it proceeds. The product replaces the manual coordination layer of a traditional CRO while keeping medical and regulatory judgment with certified oncologists and qualified staff.

Oasive builds an AI-powered research and analytics platform for fixed-income professionals, starting with agency MBS and rates. Portfolio managers, traders, analysts, and investment advisers use it to screen securities, run valuation and scenario analysis, and package investment theses, replacing the manual work of stitching together research, models, and reporting across separate tools. The platform connects each step of the workflow so a team can move from a market question to an investment decision in a single session.

Obia brokers commercial real estate leases and purchases for businesses, starting with industrial and office properties in California. Its brokers use AI agents to read listing brochures, compare rents, and check zoning across multiple properties in parallel, compressing research that once took hours per property into minutes. A shared market database grows with each search, and a local broker handles tours and lease negotiations through to signing.

ObsessionDB builds a fully managed, real-time data warehouse for teams running software and AI agents that generate data at high volume. It runs on a proprietary, more efficient fork of ClickHouse with compute-storage separation, keeping p99 query latency flat and cost per query falling as data scales from terabytes to petabytes. It replaces standard managed ClickHouse deployments where query costs balloon as data grows.

Odin Robotics builds and deploys robot arms that handle repetitive manual tasks, such as CNC machine tending, packaging, kitting, and palletizing, inside small and mid-sized factories. The robots work around existing machines and workflows, so manufacturers avoid redesigning their lines. Programmed controls handle routine movements while learned skills let the arm adapt to variations in each task, replacing the manual labor operators would otherwise perform on the floor.

Pendulum Robotics designs and manufactures custom motors and actuators for robotics and drone teams that need components built around a specific form factor and mission rather than off-the-shelf parts. Working from its Los Angeles factory, the company handles high-mix production using proprietary software and tooling, replacing the slow, generalist supply chains that hardware teams typically depend on for motion components. The target customer is a robotics or drone developer that would otherwise adapt their design to whatever standard motors the market offers.

OnePatch builds an AI site-reliability engineer that instruments a team's production stack, monitors service behavior in real time, and drives incidents from detection through to a shipped fix. It targets fast-moving engineering teams who currently lose hours correlating logs, traces, and alerts across tools whenever a deployment breaks something. OnePatch replaces the manual on-call loop: triaging alerts, gathering context, and authoring the remediation.

OpenHack builds an AI security engineer for software teams, covering codebase scanning, automated penetration testing, supply chain analysis, and secret scanning. It ingests business context, system design, and developer intent to find and verify real vulnerabilities, then filters false positives and prioritizes findings by exploitability, replacing the manual triage and testing work a human security engineer would otherwise own.

Sigma delivers a managed data lakehouse service built for companies navigating acquisitions, handling the data infrastructure work that internal engineering and finance teams would otherwise own across deal preparation and post-close integration. The service targets B2B and financial services contexts where data consolidation across entities is a recurring, resource-heavy burden.

OpEra applies mathematical optimization and AI to physical operations problems in retail, logistics, manufacturing, and agriculture. It targets VPs of Operations, supply chain leaders, plant managers, and COOs, replacing manual analysis and ad-hoc decision-making with automated, model-driven solutions. The work it displaces is the slow, labor-intensive operations research that organizations previously handled with internal teams or consultants.

Orca Aerospace builds an AI spacecraft operator that runs onboard the vehicle, replacing the ground-based human operators who currently fly each satellite through scheduled communication passes. The system decides what to do next to carry out the mission, hands execution to a deterministic planner, and filters every command through safety guardrails before it reaches the spacecraft. Satellite operators running growing constellations get continuous autonomous control between ground passes, cutting the latency and staffing load that today's console-driven ops rooms cannot scale to match.

Orgo provides cloud computers built for AI agents, giving each agent its own persistent, isolated virtual machine with a full desktop, browser, and terminal. Developer teams and product builders use it to run AI workers around the clock without provisioning or managing servers. It replaces the need to configure custom infrastructure every time a team deploys an autonomous agent.

ORO AI builds an open benchmark arena for e-commerce AI agents, running on the Bittensor network. Agent builders submit shopping agents that are evaluated daily against new, changing commerce tasks by independent validators in sandboxed environments. The strongest agent traces form a training corpus that ORO uses to develop its own model for long-horizon shopping tasks, replacing static test sets and manual evaluation of agent quality.

Papaya builds an optimization engine for AI agents already running in production. It connects to an engineering team's agents via SDK, scans production traces across prompts, context, tool calls, subagents, and models, and surfaces ranked recommendations to fix problems like context bloat, excessive tool calls, and serial API calls. Each approved fix is pushed directly to production as a pull request, replacing the manual work of reading traces and diagnosing agent inefficiencies by hand.

Particularity builds enforcement software for government consumer protection regulators. It crawls live commercial websites to detect deceptive practices, such as hidden checkout fees, fake urgency timers, obscured cancellation terms, and misleading ads, then packages the findings as ready-to-act cases. The product replaces manual investigative work that regulators would otherwise do by hand, letting agencies find violations proactively rather than waiting for consumer complaints.

Bryel builds custom multimodal AI models for businesses, trained to handle text, images, audio, video, 3D, and sensor data specific to each client's domain. Companies own the resulting models outright rather than depending on a shared general-purpose API. The product targets work that off-the-shelf frontier models handle poorly: domain-specific visual understanding, spatiotemporal reasoning, and artifact generation in the formats and standards a team already ships.

Perit.AI builds a data labeling and reinforcement learning platform for AI labs, connecting them with domain experts to produce high-quality training data. It replaces the manual, fragmented work of sourcing specialized annotators and running feedback pipelines that AI teams otherwise stitch together themselves.

Petrichor Bio builds wet lab infrastructure that lets AI agents design and execute real biological experiments and collect empirical data. The platform serves AI-driven research workflows, replacing the need for human-operated lab processes to run iterative experimental cycles. By closing the loop between AI reasoning and physical biology, Petrichor Bio enables AI systems to learn from live experimental results rather than static datasets.

Phone Harness gives AI agents direct control over iPhone and Android devices, letting them tap, type, and scroll through any app the way a human would. It targets developers building agent-based workflows who need phone automation without jailbreaks, custom APIs, or app installs. The tool replaces manual phone interaction and brittle per-app integration work.

Plek maps every role and task across an enterprise to a quantified operating model, showing headcount, cost, and AI exposure for each process. It then ranks where AI automation creates the most value and builds and deploys those AI workflows with the client's team. The product replaces manual process discovery, consultant-led interviews, and ad-hoc AI prioritization exercises that leave leaders guessing where to act first.

Ploid builds people data infrastructure for developers and companies that need reliable, up-to-date information about individuals at scale. Its index covers over two billion people drawn from professional, social, and web sources, unified through entity resolution into a single profile. Developers access it via API, CLI, or MCP to power sales tools, recruiting products, customer intelligence, identity resolution, and AI agents, replacing a patchwork of static data integrations and stale exported records.

Polemos Industries builds low-cost guided interceptors and compact launch systems for ground forces, targeting the cost imbalance where incoming threats are cheap but the weapons used to defeat them are expensive. The product lets military units afford protection in the field rather than reserving air defense for high-value fixed sites. Existing trench-level defense relies on systems priced out of reach for dispersed infantry; Polemos designs around minimizing cost per kill.

Pomerado converts websites, legacy portals, and internal tools into APIs and Model Context Protocol (MCP) integrations that AI agents can call directly. Developers describe a task, and Pomerado generates a tested, hosted integration, replacing the need for browser-based agents that re-navigate pages on every run. The target customer is any team building agentic workflows that are blocked by systems with no existing API.

Praxis builds AI agents that handle the full administrative lifecycle of a law firm, from the moment a prospective client inquires to the moment an invoice is paid. The agents run intake, conflict checks, document collection, matter tracking, and billing follow-up, while the lawyer reviews and approves each step. This replaces the paralegal and admin work that otherwise consumes most of a lawyer's day and delays client onboarding from weeks to days.

Preload builds motor-control infrastructure for physical AI systems, giving robots and autonomous machines the low-level coordination layer that translates high-level intent into precise physical action. The work it replaces is the custom, hand-engineered motor control stack that robotics teams otherwise build from scratch for each platform.

SuperRadiant builds general-purpose robotic experimentalists that combine scientific reasoning with physical action to run laboratory experiments autonomously. The systems observe, form hypotheses, and execute experiments in the real world, replacing the human researcher's hands-on trial-and-error loop across fields such as materials science, energy, and biotechnology. The target is any domain where discovery depends on physical contact with the world, not just digital analysis of existing data.

Proto builds a design tool for professional design teams that treats prototypes as the primary artifact for refining and sharing ideas, rather than a secondary step. It combines the precision of code-based design tools with the creative flexibility of visual design software. Proto replaces the back-and-forth between static mockups and separate prototyping layers by making the prototype itself the central workspace.

Quippy builds a daily social-skills training app for individuals who want to get better at high-stakes conversations. Users run short, simulated exchanges covering scenarios like pushing back on a manager or navigating a first date, then receive line-by-line feedback on what worked and what did not. The curriculum adapts to each person's recurring weak spots, replacing unguided rumination and replay with structured, private practice.

Quivly AI builds an agentic workspace for B2B post-sales teams, serving founders, customer success managers, account managers, and forward deployed engineers who manage enterprise accounts at scale. Its agents automate the work that currently lives across Slack threads, spreadsheets, and tickets: converting kickoff calls into launch plans, tracking blockers through to resolution, surfacing churn risk and expansion signals mid-contract, and assembling a complete account narrative at renewal.

RATLOG builds reusable heavy-lift rockets for commercial space launch. The vehicles are designed to carry large payloads to orbit and return hardware for reuse, replacing the single-use rockets that currently dominate heavy-lift missions.

sintem builds an AI-powered physical device for teams that runs a roster of specialized agents, covering functions like executive assistance, marketing, back-office, engineering, and research. Users delegate tasks verbally, and the agents execute them across the tools and apps the team already uses, cutting the time employees spend actively managing their phones and software. The product replaces the manual, screen-heavy coordination work that normally interrupts a workday.

Reconero builds AI-native records management software for police departments, replacing the manual paperwork that consumes a large share of an officer's shift. Its AI assistant, Alyx, reads case records, call logs, and evidence to answer natural-language queries, auto-generate reports and forms, and turn data into maps and charts. Officers get time back from administrative work and back to field duties.

Redoubt Insurance sells commercial insurance to small businesses, with a focus on contractors, trucking operators, and service providers. It handles policy placement, certificates of insurance, regulatory filings, and renewals, replacing the fragmented back-and-forth that typically falls on business owners. The agency uses automation for routine tasks and applies accumulated experience to tailor coverage to each customer's specific risk and compliance requirements.

Resin Technologies builds an AI-native electronic computer-aided design tool for PCB engineers. Starting from a set of requirements, engineers use it to research component choices, generate schematics, produce part symbols and footprints from datasheets, and route boards, replacing the manual, multi-tool workflow that spans library hunting, circuit drafting, and layout. The target user is any hardware engineer who designs printed circuit boards, from initial concept through manufacturing handoff.

Rhem Labs builds a home robot and companion hardware platform for aging adults living independently. The robot handles practical tasks, such as refilling prescriptions, booking medical appointments, ordering groceries, and arranging transportation, that family members or caregivers would otherwise manage. It also monitors health vitals, detects falls without a wearable pendant, and alerts family when something needs attention.

RightNow AI builds GPU kernel optimization infrastructure for serving open-source large language models tuned for AI agents. Its platform, RunInfra, exposes optimized model APIs with session-aware routing that keeps repeated agent calls on the same warm replica, cutting context-reload overhead. The product replaces closed-source inference APIs like OpenAI and Anthropic at lower cost, targeting developers building agentic applications.

Risklytics brokers commercial insurance for robotics, autonomous vehicles, drones, hardware, and energy infrastructure companies, replacing the slow, ill-fitted process of working with generalist brokers whose carriers have added AI and autonomy exclusions to standard liability forms. The firm reads policies directly, selects admitted and surplus lines carriers that explicitly cover autonomous and physical-world operations, and manages the full placement so that clients avoid holding uncovered risk after an incident.

River builds an AI desktop for knowledge workers: a project-based workspace where users research, write, and produce finished work alongside AI across documents, slides, sheets, and a browser in one place. It replaces the fragmented workflow of juggling files across separate apps and chat tools that lack project context. The target user is anyone doing multi-step creative or analytical work who currently stitches together output from disconnected tools.

Rodeo Robotics builds robotic technicians that perform routine vehicle maintenance, starting with oil changes, at fleet depots. The robots handle inspections, fluid services, and tire checks while human technicians review results and take over complex repairs. The system targets rental, delivery, trucking, transit, and municipal fleets that face growing maintenance backlogs and a shortage of qualified technicians.

Roma builds an AI-powered task manager for individuals who capture ideas, reminders, and to-dos on the go. It connects to a user's calendar, email, and files, then converts voice or text input into structured, actionable tasks without manual organisation. Roma replaces the friction of translating scattered thoughts into a managed workload.

Routant builds a zero-trust access layer that lets AI agents, deployed by software vendors, operate inside their customers' private systems. Vendors connect once through a single MCP endpoint, and each customer controls which apps, databases, and APIs their agents can reach, which actions require human approval, and gets a full audit trail. Routant replaces the custom access and permissions plumbing a vendor would otherwise build separately for every customer deployment.

RouterPlus runs a marketplace that routes AI inference workloads to the right model at the lowest provider price. It tests candidate models against a team's actual traffic, lets providers bid for the workload, and keeps a backup provider live for failover. Teams call everything through a single API, replacing the default practice of sending every task to the most expensive model regardless of whether a cheaper one would do the job.

Lambda Robotics builds and deploys autonomous robots for the data centers and facilities that support AI compute. Its machines handle the physical operations inside AI infrastructure, replacing manual labor in those environments. Each deployment feeds back into the system, making the robots more capable over time.

Rystic builds stateful, local API simulators that let developers test and train AI agents against realistic stand-ins for services like Slack, GitHub, Stripe, and Kalshi without hitting live endpoints. Each simulator tracks state across calls and is continuously verified against the real API, so agent behavior tested locally matches production. It replaces ad-hoc mocking, sandbox rate limits, and the risk of running agents against live third-party services during development and CI.

Saccade builds an early-detection screening tool for Alzheimer's biomarkers, targeting patients at longevity clinics before any cognitive symptoms appear. A patent-pending attachment on a standard VR headset captures eye movement patterns during a 15-minute neurologist-designed assessment, then an ML model scores Alzheimer's risk against those patterns. The tool replaces late-stage symptom-prompted diagnosis, giving clinicians a window to intervene with FDA-approved treatments while they can still slow the disease.

Sentient OS builds an on-device AI agent for Mac that reads a user's emails, messages, files, and meeting transcripts overnight, builds a local knowledge base, and prepares completed work, drafted replies, flagged renewals, promised reports, ready to send with one click. A second mode, Sidekick, takes over the active task on screen via computer use while the user moves on. All inference runs locally on Apple silicon using a custom LiteRT-LM stack, so no user data touches Sentient's servers.

Shepherd Robotics builds general-purpose robots that perform skilled physical tasks in data center construction and advanced manufacturing. The robots handle work such as cable assembly, wire harness production, kitting, and factory packaging, jobs that resist conventional automation because they demand fine dexterity and tolerance for changing conditions. Underlying the hardware is a foundation model trained on real-world and simulated data so the robots can transfer skills across different workflows without task-specific reprogramming.

Shire Intelligence builds an AI floor manager for full-service restaurants, using the CCTV cameras already installed on-site. It tracks table states in real time, recognizing when guests are ready to order, when a check is needed, or when a section is overloaded, then alerts staff and routes new parties to servers who have capacity. It replaces the judgment calls a human floor manager makes by sight, and fills in the gaps a POS cannot see because it only records what staff manually enter.

Silica builds force-sensing actuators for robots. Where conventional joints estimate force indirectly from motor current, Silica measures torque directly at the output, giving every joint real contact data. The hardware targets robotics developers working in manufacturing, lab automation, and robot training who currently write software workarounds to compensate for bodies that cannot feel.

Simantic builds a hardware simulation platform for embedded firmware development. Engineers and AI agents run, inspect, and test firmware against simulated microcontrollers and connected devices, replacing the need for physical silicon during development and validation. The platform gives both human developers and AI tooling the hardware-level visibility that was previously only available on real boards.

Simulithic builds autonomous user simulation for production monitoring. It learns real user journeys and continuously runs agent-based swarms across the actual devices, browsers, account states, and network conditions those users are on, catching broken flows, third-party failures, and regressions the moment they occur. It replaces manual QA on production and scripted end-to-end tests that miss the edge cases real users hit.

Sinter builds a full-stack plastics manufacturing pipeline for iteration-heavy hardware teams, taking them from same-day additive prototypes to small-batch injection-molded production parts on week-scale timelines. It replaces the slow, capital-intensive path of committing to expensive tooling or standing up in-house capacity, letting engineers revise designs and swap molds during active production runs.

Sintropix builds an AI-native ERP for companies operating across multiple Latin American countries. It automates bank reconciliation, multi-currency consolidation, and month-end close, replacing the manual accounting work that finance teams typically handle across spreadsheets and disconnected local tools. The platform surfaces a live financial snapshot, including cash position, burn, runway, and accounts payable and receivable, through an AI assistant called Erwin.

Sona8 builds a voice agent that conducts individual conversations with every employee about how their work actually operates, replacing the selective stakeholder interviews that consulting and transformation teams run manually today. Each worker speaks with the agent for roughly twenty minutes; the agent probes with follow-up questions and the output is process maps and structured problem lists organized by site and team. During rollouts the agent revisits employees to track whether changes landed, giving program leaders a continuously updated picture of operations across the whole organization.

Spacesheep builds a shared workspace where AI agents publish their outputs, turning a session's results into a structured, browsable page that a team can read, comment on, and refine together. It replaces the scatter of raw agent outputs dumped into chats or local files by giving every deliverable, whether a report, dashboard, deck, or internal app, a persistent, versioned URL with live data support and semantic search. Human collaborators and agents work on the same surface simultaneously: inline comments, @mentions, and an in-product agent that applies edits across every open screen in real time.

Spine runs benefits, payroll, compliance, and people ops for startups and mid-market companies, replacing both the traditional benefits broker and the PEO. Dedicated senior HR and benefits specialists operate directly inside a company's existing HRIS, whether Rippling, Gusto, Deel, or ADP, handling open enrollment, multi-state tax compliance, onboarding, and employee support. The work it displaces is the back-office HR burden that founders, finance teams, and chiefs of staff otherwise carry themselves.

Stacktrace builds observability and policy enforcement for AI agent fleets. Engineering and security teams use it to inspect every tool call an agent makes, trace which MCP server, plugin, or hook was responsible, and block or gate unapproved components before they run. It replaces manual review of agent logs and ad-hoc access controls with a continuous, policy-driven layer that fires real-time alerts into Slack or Claude Code while a session is still active.

STASIS builds smart eyeglass frames that continuously track eye health and cognitive performance throughout the day. Six or more inward-facing sensors measure over 30 metrics, including blink rate, fatigue, HRV, SpO2, and cognitive load, distilling them into a single daily eye score. The frames also auto-tint lenses and push adaptive adjustments to connected screens, replacing the guesswork consumers and athletes currently apply to eye strain, focus, and recovery.

Streamline Systems builds a drop-in navigation module for UAVs that operates without satellite positioning. The module integrates with Ardupilot, INAV, and PX4 flight controllers, delivering centimetre-precision, EW-resistant autonomous flight where GPS is unavailable or unreliable. It replaces GPS as the primary positioning layer for both commercial and military drone operators.

Super Powers Mobility (SPM) builds modular, high-voltage battery packs, power electronics, and traction powertrain systems for mobility and stationary energy storage applications. Its customers include both military programs and consumer-facing vehicle builders who need high-performance power systems but lack the in-house expertise to design safe, high-voltage packs and power export hardware. SPM replaces the fragmented approach of sourcing battery development and power electronics from separate vendors by delivering both from a single team.

Superserve builds virtual machine infrastructure for autonomous AI agents, replacing traditional cloud setups that were designed for single-app, multi-user workloads. Each agent or reinforcement-learning rollout gets its own isolated Firecracker VM that cold-starts in under 50ms, can pause with memory and files intact at no compute cost, and resumes or forks from an exact prior state. Teams building coding agents, autonomous workers, and eval or training environments use Superserve to run long-horizon tasks without managing servers, warm pools, or credential exposure inside the machine.

Sutrix builds automated data pipeline and statistical analysis software for hospital and academic clinical research teams. It ingests patient data from electronic health records, wearables, and other sources, cleans and reconciles it, and runs auditable statistical analyses, replacing the manual scripting, data wrangling, and lengthy queuing that biostatistics cores currently do by hand. The result compresses work that previously took weeks into days, freeing biostatisticians to handle more studies.

Synara builds an AI-powered remote monitoring station that works with cameras security integrators have already installed, requiring no new hardware. It watches commercial sites around the clock, detects incidents ranging from intrusions to falls, fights, and weapons, then handles the full response: verifying events in seconds, issuing live talk-downs, dispatching verified alarms to police, and coordinating on-site until the incident is closed. This replaces traditional human monitoring centers, which only watch for simple after-hours intrusions and hand every response back to the customer.

Taan.ai builds foundation music models for precise, instruction-following score generation, aimed at developers and teams creating video tools, ad platforms, and games. Given a video, the model analyzes cuts and emotional shifts, then generates an original, fully mixed soundtrack where the music changes at the exact moment the scene does. This replaces manual music supervision and the trial-and-error of current generative audio tools that cannot reliably follow timing or instrument-level direction.

Talos builds predictive maintenance hardware and software for large power transformers, assets that cost millions to replace and take years to source. Magnetic, non-invasive sensor devices attach to existing equipment in seconds, continuously capturing vibration, temperature, electromagnetic, and humidity signals, then running anomaly-detection models against each machine's learned baseline. When a fault surfaces, Talos delivers a root-cause analysis and specific remediation steps, replacing the reactive inspection routines that let problems compound until failure.

Telamont reviews construction documents, including specs, drawings, geotechnical studies, and code requirements, to flag contractual, technical, and compliance conflicts before they produce rework or legal disputes. It connects to a project's existing document storage, ingests every file and revision as it arrives, and checks each page against the full document set continuously. Construction teams replace manual document review with a tracked workflow that takes each flagged risk from detection through resolution.

Terse builds open-source durable infrastructure for AI agents and agentic systems. It replaces Cloudflare Durable Objects with a higher-memory, SQLite-native, concurrency-safe alternative that developers can run themselves. Engineering teams building software factories, multiplayer collaboration tools, and multi-agent coordination frameworks use Terse to manage persistent state across their agents.

The Agentic Data Co. designs and collects audio datasets for training speech models, serving the frontier AI labs building the next generation of voice interfaces. It records real people in real conversations, consented and licensed at the source, replacing the scraped audiobooks, podcasts, and synthetic dialogue that current models depend on. That existing data fails to capture how people actually talk in conversation, producing models that mishandle turn-taking and intent; The Agentic Data Co. collects what the open internet cannot supply at the fidelity frontier training requires.

The Company Company builds an AI agent and the supporting infrastructure to operate businesses end to end, replacing the software, workflows, and human coordination that currently run day-to-day company operations. It targets B2B customers who want a single system to handle work that would otherwise be spread across multiple tools and teams.

The Letter Company builds a growth engineering platform for software teams, combining a coding agent, IDE, and harness purpose-built for SEO, GEO, and outbound. Engineers and marketers use it to author and deploy content strategies as composable primitives, replacing ad-hoc marketing tooling and manual content workflows. The platform also tracks LLM citation rates and collects verified customer evidence, covering work that would otherwise span multiple disconnected tools and contractors.

The Subvocal Company builds Murmur, an under-chin wearable that reads jaw, tongue, and throat muscle signals to convert silent speech into text and computer commands in real time. It targets knowledge workers who need to send messages, capture notes, or control apps without typing or speaking aloud, replacing keyboard input and voice commands in situations where either is impractical. No microphone or camera is involved: the device decodes biosignals, previews actions before they fire, and lets users confirm or undo each command.

Theo builds and continuously optimizes websites for small businesses so their customers can find them on Google and in AI-powered search. An AI agent handles the full workload: constructing the site, identifying ranking opportunities, and shipping weekly updates, whether new pages, targeted posts, or technical fixes. It replaces the piecemeal work small businesses would otherwise spread across freelancers, agencies, or their own time.

Theseus Robotics builds the operating system and physical AI stack that runs humanoid robots inside supermarkets. During store hours the robots fulfill e-grocery orders; overnight they restock shelves from back-warehouse inventory. This replaces the manual picking and shelf-filling work currently done by store associates across the last stretch of retail fulfillment.