Product · AI system

Opportunity Radar — an evidence pipeline for AI-assisted research

A research system that turns a search strategy into evidence with full provenance, classifies it into signals, scores it and writes the entities into a knowledge graph — built to survive restarts.

Architect and builder
  • Go
  • Postgres
  • Durable tasks
  • Entity graph
  • LLM

The problem

Finding the right opportunities — companies, roles, signals worth acting on — means reading a lot of pages, remembering what was already seen, and judging each one consistently. Done by hand it does not scale; done by a single long AI call it is neither traceable nor repeatable.

What I did

  1. discover
  2. collect → normalise
  3. deduplicate → extract
  4. classify → score
  5. graph

A staged pipeline — discover, collect, normalise, deduplicate, extract, classify, store, graph — where every stage is a durable task row, not a function call. Each stage reads its input from the row, does one thing, and in one transaction marks itself done and queues the next, so any stage can be retried alone after a crash.

Technical detail

Provenance on every row

Evidence keeps where it came from and when, and is deduplicated across sources, so a score can always be traced back to the pages behind it.

Weights are data, not code

Scoring weights live per capability profile in the database, so changing how the radar ranks is an update, not a deploy — and a feedback loop learns from the owner's marks and explains why.

AI with a deterministic fallback

Every generative step goes through one AI layer and degrades to a deterministic fallback when no model provider is configured, so the pipeline still runs.

One knowledge graph

Companies, people and technologies are written into the same entity graph the rest of the platform uses, with entity resolution, instead of a second store.

Where it stands

Opportunity Radar runs inside the fadymondy.com platform as an internal tool, with collectors including a browser-agent source, per-profile strategies, scoring with verdicts, and a digest. It has no public app, so no usage figures are quoted.

All projects