Whatstir
Explore Y Combinator startups by sector and by the customer problems they solve.
Pick a sector, ask questions in plain English, and watch the companies each answer relies on collect in a grid.
[!NOTE]
Whatstir is an independent project. It is not affiliated with Y Combinator. Company data comes from yc-oss, an open-source mirror of the public YC directory. Read About the data before you run your own copy.
Contents
Demo video
A three-and-a-half-minute tour, from the problem Whatstir solves to how it works: the sector map, filters, groups, asking questions, profiles, compare and notes.
https://github.com/user-attachments/assets/ffe0e2e5-434d-4f02-be8f-5d640ac7faa6
Music: "Inspired" by Kevin MacLeod (incompetech.com), licensed under CC BY 4.0.
Features
- 🪴 Sector treemap. About 6,400 YC companies across 9 sectors, sized by count. Filter by batch ("Fall 2026") or show active companies only.
- 🧺 The basket. Every company in a sector as a wall of logos, narrowed by AI-suggested problem groups ("Scheduling", "Billing", …) and by name.
- 🦉 Ask, Explore, Challenge. Ask questions about the sector, a group, a batch or a hand-picked selection. Answers are written only from the companies' own descriptions and cite the companies they rely on.
- 🪢 Follow-ups that remember. Every company already discussed stays in context for the whole session, so you can just ask "Tell me more about this company".
- 🛰️ Web search when the data runs out. Questions about news, funding or anything outside the dataset get a separate, clearly labelled "From the web" part with source links, searched through OpenRouter. Companies from your data that it names join the grid too.
- 🪺 A grid that grows. Each cited company flies from the basket into a results grid that builds up over the session. The latest answer's companies are ringed.
- 🔬 Compare and profiles. Open any company for its problem card, founders and links, or compare two or three side by side.
- 🪶 Notes. Keep notes on companies or general notes, all listed on one page.
- 🌅 Daily refresh. A cron job re-reads the directory each day and enriches only new or changed companies.
/admin shows every run.
- 🗝️ Private by default. A single site password guards every page, and all AI and database calls run on the server.
Screenshots
The workspace: the company basket, the results grid (companies cited by the latest answer are ringed) and the chat.
A company profile, with its description, links and AI-generated problem card.
How it works
flowchart LR
yc[(yc-oss feed)] -->|daily cron / npm run refresh| ingest[Refresh + enrichment]
ingest -->|embeddings, problem cards, Jev group assignment| openrouter[OpenRouter]
ingest --> db[(Postgres + pgvector)]
user((You)) -->|password gate| app[Next.js app on Vercel]
app -->|retrieve candidates| db
app -->|grounded answer + citations| openrouter
- Ingest.
refresh loads the yc-oss feed into Postgres. Each company is embedded (text-embedding-3-small by default) and given a short problem card written from its own description (gpt-6-luna by default). It's also placed in AI-suggested problem groups by Jev (jev-latest). Every AI call goes through OpenRouter, with one key.
- Ask. A question is embedded, and pgvector picks the closest companies in scope (sector, group, batch or selection). Companies already discussed in the session are added back, and the whole conversation goes along. The answer model (
gpt-6.1-sol by default; any structured-output model through OpenRouter) must answer from those descriptions only and cite company ids. Citations are checked against the candidates, and the cited companies join the grid. If the model needs facts the dataset lacks, one web search through OpenRouter adds a separate, labelled part with source links.
- Keep fresh. Vercel Cron calls
/api/cron/refresh daily. Enrichment is keyed by content hashes, so unchanged companies cost nothing.
There's more depth in docs/system-architecture-document.md and docs/flow.md.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
