Machine Learning Startups funded by Y Combinator (YC) 2026

Flock Safety provides the first public safety operating system that empowers private communities and law enforcement to work together to eliminate crime. We are committed to protecting human privacy and mitigating bias in policing with the development of best-in-class technology rooted in ethical design, which unites civilians and public servants in pursuit of a safer, more equitable society. Our Safety-as-a-Service approach includes affordable devices powered by LTE and solar that can be installed anywhere. Our technology detects and captures objective details, decodes evidence in real-time and delivers investigative leads into the hands of those who matter. While safety is a serious business, we are a supportive team that is optimizing the remote experience to create strong and fun relationships even when we are physically apart. Our flock of hard-working employees thrive in a positive and inclusive environment, where a bias towards action is rewarded. Flock Safety is headquartered in Atlanta and operates nationwide. We have raised $150M in our Series E led by Tiger Global at a $3.5B valuation.

Scale accelerates the development of AI within organizations of any size to deliver critical business insights and operational efficiency. Its data-centric infrastructure platform leverages RLHF (Reinforced Learning with Human Feedback) to help organizations build the strongest AI models that supercharge their business, with customers across industries including Meta, Microsoft, U.S. Army, DoD’s Defense Innovation Unit, Open AI, General Motors, Toyota Research Institute, Brex, Instacart and Flexport.

We use AI and quantum-based simulations to discover and develop new materials for defense and dual-use applications, from energetics and propellants to carbon fibers, polymers, coatings and composites. Developing a new material today takes 10+ years of trial-and-error lab work. We compress that to months, using physics-based simulation, machine learning and AI agents that only send labs candidates that can actually be made. Sam, Ofir and Maor have spent the past two years researching the discovery of these exact materials together at the Technion lab of our CSO, Prof. Alon Grinberg Dana (former MIT postdoc), working directly with defense primes and government R&D agencies. Together we bring deep expertise in chemical engineering, computational chemistry and AI, and the production systems to put them to work. ChemPilot spun out of that work. The defense and dual-use industrial base runs on materials with fragile, sole-source and often adversarial supply chains, from energetics and propellants to carbon fibers, high-temperature composites and specialty polymers. Demand for new capabilities is surging, and materials take decades to develop. The West needs new materials it can make securely, fast and cheap - ChemPilot will help find and develop them.

Native is building an autonomous marketing platform where business can do all of their marketing within one platform using our artificial intuition model called Norn, that predicts what people will respond to, researched and built in house.

Alzheimer's testing today is bottlenecked around the onset of symptoms, when the damage is already done. Current treatments, including the FDA-approved drugs released in 2024, don't reverse the disease; they only slow it down. Earlier detection means earlier intervention. Eye movement changes linked to Alzheimer's surface sooner than any other biomarker, up to 15 years before symptoms appear. We deliver a 15-minute VR assessment delivered to patients in longevity clinics: • Tasks: patients perform visual tasks designed by neurologists to trigger a distinct oculomotor response in people with Alzheimer's or on its path • Hardware: our patent-pending attachment connects to any VR headset on the market and captures these eye movement patterns. • Analysis: the patterns are compared against our ML model to produce a risk analysis. Once classified, patients can access new interventions, such as the FDA-approved drugs. They're also referred to longevity practices, neurology practices or both, depending on their stage of decline. Testing for the earliest biomarkers allows the earliest possible intervention. In some cases, that slows the disease enough that patients outlive the onset of symptoms and never face memory loss in their lifetime.

Hundred is an automated inspection system for the components that make up consumer devices like phones, smartwatches, and laptops. Hundred helps factories cut waste and prevent line-down issues by catching sub-millimeter cosmetic defects before parts reach the assembly line. In the consumer electronics industry, designs change so fast that traditional machine vision can't keep up, forcing factories to rely on manual inspection, literally people checking parts under microscopes to sort out bad parts. Human inspectors miss roughly 15% of defects, and a single bad part can turn a finished device into scrap. Hundred is replacing manual inspection with an imaging rig that scans an entire shipping tray of parts at once. The Hundred rig blocks ambient light and uses varied controlled lighting configurations to detect defects smaller than 100µm on a range of materials. The controlled environment means Hundred only needs to scan a single good part before it can find defects, categorize them, and assess their severity accurately. Defects cost U.S. discrete manufacturers $45 billion a year, a problem Hundred's CEO, Erica, saw firsthand as a mechanical engineer designing devices at Meta, Google, and Magic Leap. She met Imani, Hundred's CPO, at Columbia University's admitted-students day in 2017 and Imani went on to spend five years at Twitch building software that handles billions of requests a day. Erica later met Iris, Hundred's CTO, in the Bay Area. Iris worked as a quant at Citadel before pivoting to study at Imperial College London where she researched machine learning on 3D meshes and other structured data. Together, they're building Hundred because manual inspection shouldn't still exist in 2026.

TrueMile is building a fully autonomous logistics company, starting with trucking. Delayed pickups, empty miles and poor planning eat into a carrier’s profits. Our agents coordinate loads, routes, driver schedules and fuel stops to earn more from each truck and lower the cost of moving goods. We’re starting with our own fleet so we control execution, learn from every delivery and keep the gains. Human drivers come first. As autonomous trucks become available on our routes, we’ll have the customers, freight and infrastructure ready for them. As we drive toward AGI, people will still need groceries, clothes and materials to build their homes. Someone has to move all of it. We want every leap in AI to show up in what people pay for everyday goods.

Tokenless is an applied AI research lab. Our goal is to end reward hacking. We do research and build tools to detect issues with environments and evals, understand model behavior in those environments, and prevent those issues from happening. Correcting them makes training more effective, makes model performance easier to interpret, and is crucial to solving the AI alignment problem. If you’re training models with RL, building environments, or running agent evals, and want more confidence in your environments, we're here to help.

Lyon builds private foundation models for banks, insurance companies, and fintechs. Each model learns from the company’s proprietary transactions, payments, clicks, and customer interactions to predict credit, fraud, collections, income, churn, and lifetime value; and runs inside the company’s cloud, so its data and intelligence remain under its control. Lyon is already working with a major insurer and trained a model on 28B transactions for a fintech serving tens of millions of active users, identifying premium-card converters 4x more precisely than its existing rules; the model is now being deployed for credit.

Dream makes AI cameras that catch damage teams miss. Car dealerships, equipment rental companies, fleets, and more move expensive assets in and out all day, and undocumented damage is a costly leak. Dream cameras are simple to set up and work automatically. Each vehicle or machine that passes is photographed from every angle, identified, and reconstructed into a complete condition record. On return or intake, Dream instantly flags new dents, scrapes, and changes your team might miss, giving one source of truth for every asset's condition over time. Ryland worked at X/xAI and built aerial damage-detection vision at Vexcel Imaging. Alexander worked at AWS and trained edge-camera models deployed in the wild across Africa and India to detect poachers.

PRINCEPS is the first purpose-built insurance company for the trillion-dollar compute buildout. We underwrite SLA, GPU residual value, and outage risk, and price the coverage that unlocks cheaper financing for data centers and neoclouds.

Mentlio an applied research lab focused on optimizing AI ROI for the modern enterprise. We help engineering teams draw a line between their AI spend and delivery, per engineer. We then cut ~30% of AI spend via Mentlio's custom ML-based intelligent routing, context optimization, and prompt compression.

Physical AI — robots and world models — is bottlenecked on one thing: diverse, real-world data at massive scale showing how humans actually do work in the physical world. It was never written down, and you can't buy it, because no one has ever captured it. Only real local experts, the people who do this work every day, can produce data this authentic and valuable. DeepReach builds the network that reaches them. We design and manufacture wearable stereo capture devices, then place them with local data partners who run collection as their own business — in warehouses, workshops, farms, kitchens and repair shops. They own the local relationships; we supply the hardware, the software platform, the quality pipeline and the payments. Anyone can run a data business. That structure is what produces diversity. Instead of thousands of hours inside one environment, we deliver hours spread across thousands of environments — which is what actually makes robot policies generalize. 475 devices in the field and 100+ data partners are live across 7 countries, adding new partners every week. They've collected nearly 150,000 clips in the past three months, with volume roughly doubling month over month. Our data is in production with multiple frontier model and robotics companies, and we're scaling to 10,000+ devices across 50 countries.

Autoresearch helps teams train agents for tasks where general models are not good enough. Give us a task, your current agent or model, and an evaluation. Autoresearch runs an autoresearch campaign across data, SFT and other post-training methods, continual learning, prompts, tools, harnesses, and systems, then returns the best verified improvement. It runs hosted or in your cloud. We provide the research system around models: persistent memory, full experiment lineage, compute orchestration, and statistical verification. We’re starting with agent and model training, continual learning, and optimization. Reach out to us for early access at founders@autoresearch.inc.

Tracer is an AI research lab building more capable and efficient AI through model coordination. We research adaptive inference, ensembling, and systems in which multiple open-source models work together to produce capabilities beyond any single model. Our work spans reasoning, code, research, and agentic systems, with the goal of making advanced intelligence more accessible and economical.

We're scaling self-improving AI scientist teams to ideate/experiment/discover, generating new knowledge autonomously. Their discoveries reduced ARC-1/2 costs by 10,000× while maintaining state-of-the-art accuracy and yielded material advances on combinatorial ML problems. We believe that AI is bounded not by compute, but ideas. Scaling to millions of sessions of discovery enables leveraging its own discoveries and presents a novel training source capable of ushering a new era of self-reliant foundational models.

Agnost AI reads every conversation a company's AI Agent has with its users to find where it's silently failing and then turns that same data into custom models that run the agent better, faster & cheaper than frontier models.

Risklytics is the insurance company for frontier tech such as physical AI, robotics, and data centers. Companies answer 6 questions and we bind a tailored policy that other insurance companies won’t.

CoArena is a live arena where anyone can use the world's top computer-use models racing two of them on the same real computer task and judging which one did it better. Every battle becomes something the AI labs can't build themselves, an honest test of their agents on real work, and the data to make them better.

We train human brain cells for compute

Capital allocation is becoming a more important part of individual economic life. Finance is moving toward a global, onchain, 24/7 future with thousands of markets trading at once. The asset management industry is built around the scarcity of financial expertise and the cost of organizing people around capital. Institutions cover global markets by employing analysts, researchers, and quants around the clock. AI turns this organizational capacity into software. Instinct is building the harness for this transition. Instinct is a proactive team of agents that searches thousands of markets, investigates opportunities, and operates 24/7. The individual becomes the portfolio manager, Instinct becomes the investment firm operating around them. As institutional advantages become widely accessible, the edge concentrates in each trader’s worldview, judgment, and bets. The first one-person investment firm managing $1 billion will emerge within two years. It will run on Instinct.

PerfectBit builds learned policies for robots that do real work. We train manipulation and control policies that hold up outside the lab, in the unstructured environments where real tasks actually happen. Founded by two alumni of Meta's Superintelligence org: Péter Vajda led Media GenAI foundation-model R&D (Emu, Movie Gen, image and video editing), and Seiji Yamamoto led teams in the Core Llama group across LLM pre-training and post-training, inference optimization, and computer vision. We're early, deliberately quiet about the details, and hiring the first researchers who will shape what this becomes. San Francisco, on-site.

Datoric develops custom datasets for voice models, robotics, and world models, treating research, collection, verification, and production as one continuous process. We work closely with frontier model teams to turn emerging limitations and model failures into testable data hypotheses, while running our own experiments ahead of customer demand. This allows us to operationalize validated methods into repeatable collection systems at scale. Data is collected through private invite-only applications separated by modality, customer, and trust level. Every submission remains linked to the contributor, device, task, session, consent, rights, and processing history that produced it, giving our internal QA and fraud models the context to detect problems that may appear legitimate in the finished file. Each collection reveals new failure cases and quality signals that improve the systems behind the next dataset. Once a collection method is validated, we scale it up as a solution to the model failure and it also becomes a reusable data recipe for future custom projects or independently collected, rights-cleared data products.

Velum is the operating system for data quality. Velum automatically monitors and enforces data quality across a company's data stack, so bad data never reaches dashboards. We turn data quality from a manual task into infrastructure that runs itself. Data trust you can prove. From the pipeline to the boardroom.

Darwin Robotics creates general-purpose robots to build and maintain the machines civilization runs on. We're starting with AI infrastructure, where skilled technicians perform a long tail of physical work across assembly, installation, commissioning, and maintenance. Our robots learn new tasks from demonstrations rather than being explicitly programmed for each workflow. First deployments are planned for early 2027.

Allus builds next-gen vision foundation models that bring real intelligence to manufacturing. Enabling factories to see, understand, and improve production in real time.

Hyperspell is your Company Brain. AI agents are brilliant and clueless. They ace any test and still have no idea how your company works. Hyperspell connects your tools and synthesizes documents and conversations into a live, permissioned context graph. Any agent can read from it and write back to it like a filesystem, with every fact traced to source.

Amika is infra to build your software factory. We give your team sandboxed AI coding agents in the cloud: the same infrastructure Ramp, Coinbase, and Stripe built in-house. Kick off work from a web UI, CLI, Slack, or API. Use your preferred coding agent (Claude Code, Codex, etc.). Each agent understands your codebase, runs autonomously, and ships real PRs with live previews of the apps it changed. Use it as your daily driver or plug it into your own code-gen pipeline via API.

Brickanta – agentic AI for society builders. Hundreds of construction-specific AI agents for project analysis, insights, tenders, procurement, and more. Combine the latest AI technology with your organization's project data, templates, and workflows to identify risks and opportunities early and produce better decision support. Brickanta has raised $8 million from leading AI, construction, and real estate investors behind companies like OpenAI/ChatGPT, Airbnb, Klarna, Spotify, and Plangrid-Autodesk, as well as star investors such as Mario Götze, Anton Osika, and Northzone (see movie and press). The founding team has been building and implementing AI in construction and industry since 2018 at companies such as ABB, Fabege, Husqvarna, IKEA and Konecranes.

Bezel helps fashion brands create virtual photo/video shoots with AI. Upload images of clothes, select the human you want to model it, and Bezel generates pictures and videos that rival a full studio production. Every detail of the clothing is rendered flawlessly. Try it for yourself.

OnDeck is the infrastructure layer that makes Vision Language Models accessible and scalable for enterprise. Our model, Perception-0, let organizations instantly find any object, behavior or event, in any footage, without needing to train a model or collect any training data. The Pain: Creating vision models usually takes months: collecting training data, training, then deployment. Worse yet: + it’s often impossible to get enough data for a specific task, and + even the best cv models struggle to generalize across diverse camera setups, workflows and environments. To overcome these blockers, we bet early on the power of VLMs and built a vision engine that can generalize across any task and doesn’t need any training data. We published a NeurIPS workshop paper showing our new methods with VLMs beat traditional CV even at niche tasks. Now, our model Perception-0 beats the best VLMs including Gemini at long video understanding. Our current customers include: - National Defense Organizations - Robotics Research - Security cameras - Behaviour analysis for port monitoring - Off-shore oil & gas monitoring

We are a full-service sell-side residential brokerage that lists and markets your properties. We also screen and schedule tenants showings with our AI agent for a third of the normal commission.

Lilac sells flexible GPU contracts and runs the software layer on top: monitoring, bare-metal provisioning, and managed Kubernetes and Slurm. We also run this layer for GPU cloud providers as their named operations and support partner.

DeepAware (Robotics Center of Silicon Valley, https://roboticscenter.ai/) is the fastest way for enterprises and researchers to get robots and robotics parts in the US — 72-hour delivery or Bay Area pickup. Beyond hardware, we help teams collect teleoperation data, build reinforcement learning environments, and deploy robots into production. Customers include AI labs, industrial operations, research teams, and event producers.

Kairos closes the last-mile reliability gap in AI deployments. We bring frontier techniques to companies in critical industries, deploying specialized agents that encode operator expertise and reliably automate their most manual workflows.

Kashikoi is a simulation engine to benchmark AI agents. We generate CPU friendly world models that autonomously interview agents and generate deep behavioral assessments. We built a similar technology at Moveworks which was used to ship 250+ enterprise agents to customers daily.

Photonium is building software to automate optical system design. We supercharge optics experts with intelligent tooling to reduce costs and deliver faster. We handle the full design stack — from optimization, verification, sourcing, to prototyping — for AR/VR, quantum, biotech, metrology/chip fab, LiDAR, and more.

Theorem is training models that make program verification 10,000 times faster. Using verification as a feedback loop, developers have found zero-days in GPU accelerated code and cryptography implementations, and sped up code migration in legacy systems. If you have complicated code that needs to be correct and secure, sign up for our beta!

A fully open-source, performant and actionable ML model training platform

Plexe builds predictive ML models from a problem description. It connects to data sources, conducts experiments, evaluates and deploys the models to an API endpoint.

Mundo develops data, evaluations, and applied research for frontier labs and AI companies to advance perceptual intelligence across audio, video, and emerging modalities.

Nitrode is a research company focused on improving game development with AI. To advance game development, we believe AI models firstly need to be judged against robust evaluation frameworks that reflect the actual complexity of the field.

Osmosis is a post-training platform that helps companies fine-tune language models using reinforcement learning. We work with fast-growing AI companies to train task/domain-specific models that beat foundation models on performance, cost, and latency. Our platform handles compute orchestration, reward modeling, and training run observability as a CLI-based product usable by developers and agents.

Mecha Health builds foundation models to automate x-ray analysis for radiologists. We take medical images and process them using proprietary models to produce accurate draft medical reports. Our first model was built in less than two months, and beat Microsoft, Google, and OpenAI on clinical accuracy metrics. On top of that, it’s two orders of magnitude smaller and trained with a quarter of the data. We are partnering with the largest privately owned radiology practice in the US and a multinational tele-radiology company to provide them with their own foundation model, enabling their radiologists to go from reading 1 scan per hour to 1 scan every 5 minutes. By charging on a per scan basis, x-ray report generation represents a 40B+ market opportunity.

Metreecs helps retailers plan, buy, and allocate products using AI-demand forecasting. We prevent overstock and out-of-stock situations, allowing clients to eliminate waste, free up capital, and drive higher sales.

We're building Matcha, the faster, smarter way to hire clinical talent—helping hospitals find active, qualified candidates at a fraction of the time and cost of job boards or headhunters. Matcha engages candidates at scale, matching them to your organization by qualifications and cultural fit. Recruiters save time, hospitals save money, and candidates get a better experience—a radically better hiring model for everyone involved.

Archil transforms S3 buckets into a 30x faster, unlimited, local disk. Archil enables AI, analytics, and serverless applications to instantly access massive data sets without waiting for data transfer. Researchers use Archil for shareable, local storage of data set and model versions that never runs out of capacity.