Knownworld

Years of chats hide a hundred open doors. Knownworld maps the people you actually know, finds out who they became, and turns that into warm intros to jobs, partners and the right person in any city.

Comment

Demo

Application: https://knownworld-web-ncr73a6xhq-uc.a.run.app

File to Test (Chats): https://github.com/OGcryptonaut/knownworld/blob/main/sample-data/result.json

Codebase

https://github.com/OGcryptonaut/knownworld

Inspiration

Everyone tells you your network is your best way into anything: a job, a partnership, the right person in a new city. Nobody tells you how to actually query it.

I'm Vlad. Seven years in Web3, hundreds of Telegram contacts, my whole professional life in one messenger. As data it's useless: an 878 MB JSON export, names I half-remember, companies people left years ago. So when I need someone specific, I scroll. I send the same message twenty times. People ignore me :) People changed jobs and I missed it. And someone I actually know is already inside the company I'm knocking on. I just don't know it.

So my friend Sahil and I built Knownworld: your chat history, turned into a private, enriched contact database you can ask questions.

What it does

  • Ingest. Your Telegram export parses entirely in the browser. Raw chats never reach a server.
  • Distill. Gemini reads chats in transient ~20-chat batches and returns strict-schema contact rows: who they are, where they work, a two-line summary. Closeness is computed in code from volume and recency. A model never scores it.
  • Research. One grounded lookup per contact (the query is a name plus a company, nothing else): what they do now, how they can help you, work history, location on a map, citations on every claim. The match / mismatch verdict is computed in code by comparing evidence to the database.
  • Requests. You just ask. "Is there a BD or partnerships job for me, posted in the last 30 days?" hits the live public ATS feeds of your contacts' companies and returns real postings, each with its warm path: the people you know there, ranked by closeness. "Who should I meet at an AI conference in San Francisco?" ranks your actual people, with grounded reasons and sources.

The app never sends messages anywhere. Drafts are copy-out only, you send them yourself.

How we built it

Two Cloud Run services. The agents service is FastAPI + Google ADK: refine, enrich + verify (Google Search grounding), a planner / matcher / web scout / composer behind the chat, and a job scout with five ATS clients. Gemini 3.5 Flash via Vertex AI, structured output on every call; malformed output is rejected with reasons, never patched. Firestore holds the distilled rows, one isolated tenant per account. Research fans out through Cloud Tasks. Secrets live in Secret Manager. The dashboard is Next.js.

Self-deploy is the product: one script stands the whole stack up in your own GCP project. No shared server, no operator who can read your data.

Challenges we ran into

  • Nobody sane uploads a decade of DMs to a hackathon project. So the privacy boundary is architecture, not promises: parse in the browser, transient batches, only distilled rows persist.
  • Models love to overreach. Ours tried to output closeness; the schema has no field for it, and a hostile fake model in our tests proves the value gets dropped. Hallucinated contact ids are dropped in code. Thin evidence stays "unverified", never a guess.

Accomplishments we're proud of

The full loop runs live: the 15-contact demo distills in one batch, all 15 research cards come back with code-computed verdicts and citations, and a jobs request scans thousands of real current postings across 9 live feeds with warm paths on each.

What we learned

That we need to build it for ourself it really works in terms of finding the right persons.

What's next for Knownworld

A fully local-first variant (no cloud at all), more messengers, and richer research passes. The repo is open under MIT, the demo network of 15 famous founders is included, and the live instance is one click away. Tell us what breaks.

Built With

Share this project:

Updates

Submission history