FeatureRadar

About the project

A pipeline that reads what people complain about in fintech, pulls out real problems, groups similar ones, researches who already solves them, and writes scored business ideas.

FeatureRadar looks for business ideas in what people complain about. It reads fintech discussions on Reddit, Hacker News and consumer complaint sites, pulls out real problems, groups similar ones, researches who already solves them, and writes scored ideas with an MVP and a launch plan.

The key design choice was evidence: every idea can be traced back to the exact posts it came from. It ran locally and produced 15 ideas before I stopped in December 2025.

How it works

1. 1

   Collect

   Scrape posts. An AI filter scores each one from 0 to 10 before it's saved.

2. 2

   Sort

   Each post is tagged as frustration, missing tool, willing to pay, or wasted time.

3. 3

   Extract

   Pull out the problem, who has it, their workaround, quotes and how badly they want a fix.

4. 4

   Group

   Similar problems are clustered by meaning and each group is scored.

5. 5

   Research and write

   One agent checks the competition; another writes and scores the ideas.

Under the hood

Pipeline

TypeScript services, models and embeddings through OpenRouter.

Data

Postgres with Prisma in Docker, with every agent step saved.

Sources

Reddit, the Hacker News API, Firecrawl and Apify.

What I did

• Replaced the first scaffold with my own codebase, then rebuilt version 1 as a 5-step pipeline that keeps the evidence for every idea
• Step 1 sorts posts into 4 signals: frustration, missing tool, willing to pay, wasted time
• Step 2 pulls out the problem, who has it, their workaround, quotes and how badly they want it
• Step 3 groups similar problems with embeddings, and scores each group
• Step 4 is an agent that researches competitors with search tools and saves every step
• Step 5 writes ideas scored on practicality, revenue, demand, competition and novelty, with an MVP and a launch plan
• Scrapers for Reddit, Hacker News, consumer complaints and more, with an AI filter before saving
• An API and a small page to browse ideas and the research behind them

What was new

• Following any idea back to the exact posts it came from
• Where you scrape matters more than the prompt: the Hacker News front page gave 0% useful signal, GitHub issues gave 63%

Problems I hit, and how I fixed them

1. 01

   Version 1 produced 2,000+ weak ideas.

   Fix The 5-step pipeline filters noise before any idea is written.

2. 02

   The AI's JSON broke in many ways.

   Fix Strict output schemas plus cleanup of the replies.

3. 03

   About 70% of problems matched nothing and were thrown away.

   Fix Give unmatched problems their own score.

4. 04

   Only 9% of scraped posts held a real problem.

   Fix An AI pre-filter that scores each post before saving.

Screenshots

To be clear

Several research tools are still stubs, and in the last run no problems grouped together, so each idea rests on a single post. The brief and first scaffold came from the repo's owner.