All projects
Market RadarInternal tool

Deciding what to build next, with data

Market Radar is Jednorog's internal research tool. It collects public market data every day and shows which niches are emerging, accelerating or getting crowded.

Role
Founder · built solo
Since
2026
The problem

A hunch is a bad market model

Picking the next product usually comes down to a hunch and a few loud posts. The useful signals, such as launches, hiring, search demand, funding filings and developer activity, are public but scattered across dozens of sources nobody checks every day.

How it works

One market thesis, checked every day.

AI made software cheap to build, so the hard part moved to distribution and trust. Market Radar tests that thesis against 30+ public data sources every day, grades eight hypotheses and ranks startup niches by demand, competition and stage, to decide what Jednorog builds next.

  1. 01

    Collect

    Scheduled collectors pull launches, repositories, package downloads, filings, hiring and search data into an append-only history. Every value keeps its source and date.

  2. 02

    Score

    Signals vote on eight hypotheses about the market, and each niche gets a stage: emerging, accelerating, crowding or consolidating.

  3. 03

    Review

    Every Monday an AI agent writes an analyst memo with forecasts. It can only cite stored facts, and the server rejects any claim it can't trace.

What it does

Inside the product

  • A hypothesis scorecard that the owner signs off
  • Opportunity map: demand, market size, competition and regulation
  • History back to earlier hype waves for comparison
  • Sentiment panel across engineers, founders, investors and prediction markets
  • TV board mode for a screen on the wall
  • A private problem diary fed by recurring complaints online
Status

Where it stands

  • Running in production as an internal tool
  • More than a dozen build phases shipped in its first weeks
  • Not public: it feeds Jednorog's own decisions

Role: Founder · built solo

Under the hood

Engineering notes

Stack

  • Bun
  • Hono + oRPC
  • Postgres
  • React 19
  • Claude Code
  1. 01

    The AI memo can't type numbers. It cites facts by ID, and the server drops claims that point to nothing real.

  2. 02

    The agent runs as a scheduled Claude Code routine and talks to the app only through a data branch on GitHub.

  3. 03

    History is append-only: corrections are new rows that supersede old ones, each with a source link.

  4. 04

    Forecasts go into a ledger and are scored when they resolve.

Want to talk about Market Radar?

A partnership, an investment, a pilot or just a question. Write, and I'll reply personally.

Get in touch