Engine Architecture
Seven stages. Zero guessing.
Every opportunity brief is the output of a deterministic pipeline — not an AI hallucination. Here is exactly how it works.
Ingest
Reddit, Hacker News, GitHub Issues, StackOverflow
The engine reads ~10,000 threads per week across four platforms. Every post is normalized, hashed, and deduplicated before touching the pipeline.
Extract
Gemini identifies pain signals
Each post is passed through a Gemini prompt that extracts the underlying pain description, buyer type, and workaround mentions. Non-signal content is discarded.
Cluster
Semantic grouping across sources
Signals describing the same pain in different words are grouped into clusters. A complaint on Reddit and a GitHub issue about the same broken workflow become one cluster.
Score
Five-dimension base score + overlay metrics
Each cluster is scored on pain acuity, frequency, willingness to pay, search demand, and MVP simplicity. Overlay metrics add signal confidence, temporal slope, and graveyard risk.
Promote
Only high-conviction clusters advance
Clusters with a base score above the promotion threshold and at least 3 distinct signals are promoted to opportunities. Everything else enters the graveyard archive with a documented failure pattern.
Monitor
Velocity tracking across runs
Each cluster is tracked week-over-week. Accelerating evidence velocity signals an opening window. Flat or declining velocity signals narrowing or closing.
Project
Safe read-model for subscribers
Promoted opportunities are projected into two layers: a public Signal layer for all users, and a locked Intelligence layer with temporal models, graveyard patterns, and structural asymmetries.
See it in action.
The engine runs every week. Browse this week's validated opportunities — each one is the output of this exact pipeline.
Browse Opportunities