GTM engineering
Signal Scout
Reads founder groups on Telegram, classifies every post with Claude, and alerts me in Slack before anyone else sees it. Quality over volume: it stays quiet unless the lead is one I would act on. It logs the lead in Notion, and a person approves anything before it goes out.
I built this demo as one self-contained page: no video, no backend. Companies, contacts and numbers are fictional; the mechanics are real.
Best viewed on a laptop: Open the full demo
Context
I was spending 5 to 10 hours a week scrolling WhatsApp, Telegram and Slack groups for leads. It ate the time, and when I reached a post 24 hours after it went up, the lead had already gone cold. I built Signal Scout to do the reading for me: it finds the posts that match the ICP, so I message the lead while it's still warm and get those hours back. Volume was never the point: one alert I would act on beats a hundred maybes, so it is built to stay quiet when in doubt. For me, the ROI has been significant.
What I built
- Listens 24/7 to allowlisted founder and tech groups on Telegram, through a read-only session on a dedicated number
- Logs every message first, so nothing is processed twice or lost
- A keyword prefilter drops the obvious noise before any model call
- A two-stage Claude classifier: the model extracts the facts (who posted, company, roles, ICP signals), then a fixed rule makes the call, never the model alone
- A Slack alert in under a minute, with the company, the roles and the reason it fit. When in doubt it stays quiet
- Logs the lead in Notion and adds a first-message draft, built from the post and public company facts only (no LinkedIn, no research on people)
- A person approves the draft. The system itself never posts or sends, on any platform
How it runs
Deployed via PM2 with a boot-time scheduled task. A secret-scan GitHub Action runs on every push. A two-day silent classifier outage has its own root-cause writeup: operational evidence of what failed and how it was diagnosed, alongside the fixed decision rules and human approval.
What it taught me
In GTM engineering, timing is most of the value: the same lead is warm in the first hour and cold a day later, and only a machine reads the market that fast. As a builder, I learned to build into the stack the team already lives in (Slack for the alert, Notion or any CRM for the record, a person for the final yes) and to let the AI do the reading while a fixed rule makes the decision.
Stack
- TypeScript
- Telegram MTProto (read-only)*
- a dedicated phone number
- Claude
- Slack
- Notion*
- PM2
* Swappable, and the easy part: more sources (WhatsApp and Slack would be new adapters, not built yet) and HubSpot or any other CRM can take Notion's place.
Want the walkthrough?
The code is private. The walkthrough isn't.
Book 30 minoranov.con@gmail.com