The Specialists Arrive
Hey there, investors and innovators!
The general-purpose assistant is no longer the interesting part of the AI story. This week's picks each choose one job and go deep on it: building a game, trading a market, drafting a contract, proving a piece of code correct, answering a plumber's phone. All five teams are between two and five people, and the founders tend to come with a competitive or operating edge in the domain they're attacking. What stands out more than the ideas is how soon money shows up. Several of these companies are reporting revenue, contracts, or fast user growth within weeks of launching, which says something about how ready buyers are for narrow tools that do a defined piece of work.
What the Pattern Says
- Specialist beats generalist. A market-native trading model, a legal agent scoped to contracts and corporate work, a verification stack built on Lean, and a voice agent for home services all start from one narrow, well-defined job. The bet is that depth in a domain, and the data that comes with it, wins over a general model asked to do everything.
- Early revenue is arriving fast. Several teams report meaningful numbers within weeks of launch: a 40% month-over-month revenue gain with tens of thousands of users, six-figure contracts and an LOI, and MRR growing at 20% a week. All of it is company-reported and early, but the pace suggests buyers are ready to pay for a tool that does a bounded piece of work.
- AI labs are becoming customers, not just competitors. Formal verification data and reinforcement-learning environments are products that frontier labs buy to train and evaluate models. A company selling picks and shovels to the labs has a different risk profile than one competing with them, and it points to where training data demand is heading.
- The outcome is the product, with a human behind it. An AI that handles the legal work and brings in vetted lawyers for hard cases, and a receptionist that sends emergencies to the owner, both promise a result and build in a person for the exceptions. Pricing follows: a flat monthly fee, or a quote set against what a firm would charge.
- Competitive-pedigree founders are showing up across the batch. Olympiad and programming-contest finalists lead several of these teams, from informatics golds to ICPC finalists, and one of them has built a community of roughly 12,000 contest medalists. Hard, measurable competition is proving to be a useful filter for the kind of technical depth these narrow, research-heavy products need.
Eight Capital tracks every YC batch, cohort to cohort. Reach out if you'd like a deeper cut on any theme or company above.
Fund Manager Picks
Summer Engine
The AI game engine.
Why it's interesting
(San Francisco, 3 people): An AI game engine where people and agents make, play, and sell games. Users describe a game in chat, the engine builds it, and one click launches it for multiplayer, with publishing to Steam, iOS, Android, and the web. It also works alongside Claude Code, Codex, and Cursor. The CEO started and exited his first company at 19, and the CTO worked on AI at roger.ai (acquired by FLEETCOR) at 19 before leaving a PhD to build this. The company reports 60,000+ users, all organic, with revenue up 40% over the last 30 days and a free library of 240,000+ community-made assets (company-reported).
Mnemonic Labs
OpenAI for Financial Markets
Why it's interesting
(San Francisco, 2 people): A "neolab" training a foundation model for quant trading that learns directly from market activity rather than from text, on the thesis that markets are a spatial, dynamic domain where specialized representations beat language models. The company says it built its training pipeline and architecture in under six weeks and that its current run uses over 25TB of tokens. The founders are brothers who both studied math at MIT: the CEO was a quant researcher at Citadel and Walleye, and the Chief Scientist was a quant at Tower Research Capital. No customers or revenue are listed yet.
Index Law
Your company's AI lawyer
Why it's interesting
(San Francisco, 5 people): An AI legal agent for businesses, built to do the legal work itself rather than speed up a lawyer's day. Its model, Mike-1, reviews and negotiates contracts, handles corporate matters, keeps company documents and history in one workspace, and brings in vetted human lawyers when specialist expertise is needed. The CEO is a repeat founder who started an AI education company in high school. The company reports contract turnaround under 60 minutes and quotes up to 80% below traditional law firms (company-reported), and notes that Mike-1 is not a licensed or practicing lawyer.
Repovive
Scaling AI formal verification for chips and critical software
Why it's interesting
(San Francisco, 5 people): Formal verification data and reinforcement-learning environments in Lean for AI labs, with a longer-term goal of a "black box for formal verification" for chips and critical software, where you state what to prove and the system returns an independently checkable proof. The CEO is a CS PhD candidate, a national informatics olympiad gold medalist and ICPC World Finalist who has published at NeurIPS and ICLR, and the CPO is a former Google and YouTube engineer who led teams on Google Home. The company reports roughly $100K in contracts across two environments, closed in the past month, plus a $200K letter of intent for broader coding data (company-reported).
Rezora IO
Shopify for AI voice agents: home services live in 10 minutes
Why it's interesting
(New York City, 5 people): "Shopify for AI voice agents," starting with AI receptionists for HVAC, plumbing, roofing, and other home service businesses. The agent answers questions using the business's own prices and policies, qualifies jobs, books them on the calendar, and routes emergencies like burst pipes straight to the owner. Setup takes about 10 minutes and the plan is $589 a month. The company reports 40+ businesses live, 20% week-over-week growth since launch, and $23K in monthly recurring revenue as of November 17 (company-reported).

