SAT and ACT prep is a $2 billion industry built on decades-old pedagogy. Miyagi Labs wants to replace the whole stack with an AI tutor that actually adapts to how each student learns.
Moving data between Salesforce, NetSuite, and SAP is the kind of work that makes people quit their jobs. Agentin AI built agents that handle it, and they use reinforcement learning to get better when things go wrong.
Most ed-tech products are built by people who have never taught a class. Alice's CEO spent eleven years in university classrooms, and it shows. 100K students, $20K MRR, and a product that turns your messy PDFs into actual exam prep.
Upload a photo. Add a voice. Get a real-time AI avatar you can have actual conversations with. Dollyglot is building Character.AI with faces, and the tech underneath is genuinely impressive.
Two teenage founders built an AI grading tool to $200K ARR before they could vote. Now they're scaling it across Australian schools with Y Combinator backing.
Students are already using ChatGPT to cheat on homework. Excellence Learning wants to give them an AI that actually teaches instead.
Grading STEM homework is tedious, time-consuming, and one of the main reasons adjunct professors question their life choices. GradeWiz automates the entire process, handling math notation, code, and handwritten diagrams while providing step-by-step feedback.
Every expert, creator, and educator hits the same wall: there's only one of you, and everyone wants your time. Delphi built a platform where you can create an AI version of yourself that thinks, talks, and responds the way you do.
Another AI email tool, yes — but the design decision to live inside your existing inbox instead of replacing it might be the only one that matters.
Most AI agents break and stay broken. Giga ML builds agents that learn from their failures in real time and get better without human intervention.
AI tutoring for K-12 math that actually follows the curriculum your kid is already using in class. Not a homework cheat tool. Not a calculator with a chatbot stapled on. A tutor.
AI agents are great at short tasks and terrible at long ones. Aviro built a reinforcement learning layer that teaches agents to stop breaking down after step twelve.
Deep Personality wants to compress years of self-knowledge into a single sitting, and the uncomfortable part is that it might not be entirely wrong.
The pitch is simple: what if your community platform wasn't someone else's product?
Formal verification has always been powerful and almost completely impractical. Theorem is training models to make it 10,000 times faster, and the early results on GPU code and cryptography are hard to ignore.
The pitch is seductive: just say what you want and the AI figures out the rest. Whether that holds up past the demo is a different question.
Everyone I know who's tried to grind DSA has the same story: two weeks of enthusiasm, one hard graph problem, and then the tab quietly closes forever.
Describe the prediction you want to make, connect your data, and Plexe builds, evaluates, and deploys the model for you. That is either the future of ML or a very good demo.
Everyone knows RLHF makes models better. Almost nobody has the infrastructure to do it themselves. RunRL is betting that turning reinforcement learning into a managed service is the right abstraction.
Industrial robots cost six figures and need expert programmers. The Robot Learning Company thinks dual-arm robots under $10,000 with self-learning software can change that math entirely.
Designing optical systems still involves expensive software, manual iteration, and a lot of guesswork. Photonium wants to automate the entire stack from design to prototyping, and their team has the physics chops to pull it off.
Everyone uses ChatGPT to write now. Almost nobody is happy with how it sounds. TypeOS is betting that the real product isn't better generation but better translation from AI-speak to your voice, and 6,000 users seem to agree.
Every AI team knows they need better evals. Most of them are still building annotation pipelines by hand, burning weeks on infrastructure that has nothing to do with their actual product. MangoDesk generates custom eval and post-training data pipelines in seconds. That is either a huge time-saver or a huge shortcut, and the difference matters.
Notion and Google Docs help you write things down. ChatGPT helps you ask questions. Opennote combines both into a notebook where an AI tutor named Galileo reads your notes and helps you actually learn from them. Over 55,000 students are already using it.
Every company needs explainer videos. Almost nobody can afford to make them well. Knowlify wants to turn any document into a polished animated explainer in seconds, not weeks, and the 3Blue1Brown comparison is not accidental.
Most AI companies want bigger models on bigger servers. Stellon Labs is going the other direction, building frontier-quality models small enough to run on phones, wearables, and embedded devices. Their open-source KittenTTS hit 8,000 GitHub stars in two weeks.
Everyone wants to build AI agents that use computers like humans. The problem is you cannot train those agents on production systems without breaking things. Halluminate builds realistic sandbox environments that replicate Salesforce, Slack, and enterprise tools so AI labs can train and benchmark computer use agents safely.
Everyone is fighting over which model to run. Luminal is fighting over how fast you can run any of them. Their ahead-of-time compiler turns AI models into optimized GPU code and is already beating vLLM and TensorRT-LLM on throughput benchmarks. Three people, $5.3 million, and a very different theory of how inference should work.
AI agents are shipping into production with no way to safely test them against real-world failures. Playgent builds high-fidelity sandbox environments where teams can reproduce bugs, run evaluations, and train agents through reinforcement learning, all without touching live systems.
Teachers spend five to ten hours a week grading math assignments by hand. Frizzle processes entire class sets in minutes, reading handwritten student work and generating step-by-step feedback that actually helps kids learn.
Most companies running GPU clusters have 30 to 50 percent of their capacity sitting idle at any given time. Lilac unifies scattered GPU resources across clouds and on-premise into a single compute fabric, letting data scientists submit jobs to one scheduler instead of hunting for available machines.
AI agents fail in ways that are hard to reproduce and harder to diagnose. Fulcrum builds red-teaming agents that find the failures for you, uncover environment bugs, and catch reward hacking before it ships.
Language apps have spent a decade gamifying vocabulary drills. Pingo AI skips the flashcards and drops you straight into conversation with an AI tutor that actually talks back.
Everyone is building AI agents. Almost nobody can test them properly before deployment. AgentHub provides the simulation environments, tracing tools, and evaluation frameworks that agent developers need.
Nobody reads the documentation. Nobody watches the full training video. Golpo converts your PDFs and docs into short whiteboard explainer videos with AI narration, betting that animation and a human voice can make boring content actually consumable.
Voice AI is eating the world, but the training data supply chain is still held together with duct tape and crowdsourced recordings of questionable provenance. Panels wants to fix that with 10,000 vetted contributors across 100+ countries.
AI models that generate speech and video need training data that sounds and looks like real humans. Liva AI collects it with consent, diversity, and production quality baked in from the start.
Every AI company has a GPU performance problem. Most solve it by hiring expensive kernel engineers. Wafer automates the entire workflow, from profiling to patching, and claims 1.5 to 5x speedups without anyone touching CUDA.
Every parent wants their kid to spend less time gaming. Aleph Lab is betting on the opposite: what if the gaming itself was the classroom? Their AI buddy teaches kids English inside Minecraft, and the linguistics research says it might actually work.
Building ML models still requires a team of specialists, months of iteration, and a GPU budget that would make most startups cry. Thesis thinks AI agents can compress that entire process into something one researcher can do in an afternoon.
Duolingo turned language learning into a game. Parrot wants to turn it into a feed. Short-form video content built on Comprehensible Input methodology, with power users spending 6 hours a week on the platform. The founders tested 52 language apps before building their own.
LiveBench releases new benchmark questions every month so AI models cannot memorize the answers. Eighteen tasks across six categories, all with verifiable ground-truth answers. No LLM judges. No stale data.
Doomersion turns the worst habit of the 2020s into a language learning method. Watch TikToks in your target language, matched to your level, and get smarter while you waste time.
Training robot policies in the real world is slow, expensive, and bottlenecked by hardware. One Robot builds simulations that look real and behave real so you can iterate without the robot.
Booko predicts which appointment slots will not fill and automatically offers targeted discounts to maximize utilization, with early customers seeing a 20% revenue uplift.
AxionOrbital builds foundation models that translate raw radar satellite data into optical-quality imagery, enabling continuous visibility through clouds and darkness.
Rubric AI builds the human and computational layer that converts expert judgment into training signals for frontier models, focusing on regulated industries like healthcare.
Carrot Labs builds specialized LLMs for your business workflows, then continuously improves them against your success metrics, capturing proprietary know-how that gets more valuable over time.
Velum Labs monitors production query traffic to detect when teams calculate the same metric differently, traces root causes through query lineage, and generates enforceable data contracts to prevent future breaks.
Remy AI builds warehouse robots that automate operationally intensive, high-dexterity tasks using a proprietary model that learns new tasks with less training data than standard approaches.
Polymath builds complex multi-tool environments for training AI agents through reinforcement learning. Their Horizon-SWE benchmark tests what happens when you ask AI to do real work across real tools over long time horizons.
Librar Labs is building self-healing data infrastructure that turns fragmented library records into agent-ready systems. Their school library product reportedly doubles student reading rates.
Most AI training data has clear right answers. Traverse focuses on the hard part: capturing how humans make subjective decisions in law, healthcare, sales, and writing to train models that need judgment.
Lamina Labs builds Pictor, an AI video generation tool that creates mathematically precise educational animations from a text prompt in under eight seconds, with zero hallucinations and zero approximations.
