builder-log Public logbook of how I'm building a demand intelligence pipeline — a system that surfaces monetizable pain points from online conversations and turns them into actionable product opportunities.
The pipeline (two stages) Stage 1 — Data collection & pre-filtering
Source: Hacker News Ask Stories (Firebase API, no authentication, no rate limits)
Fetches posts with full comment threads (10 comments per post)
Merges post body and comments into a single content block for analysis
Output: raw_posts.json
Stage 2 — Local LLM deep analysis
Runs on Qwen 7B via Ollama (zero API cost, completely offline)
Extracts from each post:
3 concrete, specific pain points
Scores: pain intensity, urgency, willingness to pay (1–10 each)
Product adaptability (0/1)
Product title suggestion
One-line selling point (copy-ready for Gumroad/Twitter)
Decision: IGNORE / BUILD (adaptability = 0/1)
Key insight added today: High pain score ≠ high product fit. A post can score 31/40 (high pain) but be unadaptable (complaint about a proprietary service), while a 26/40 post (decision fatigue between tools) is perfectly adaptable. I'm adding a "solutionizability" dimension to future prompts.
What's working End-to-end pipeline from HN Ask → scored JSON in ~3 minutes (manual trigger)
100% comment capture rate (15/15 posts today)
Copy‑ready output (pain points + product title + selling point)
Zero external API costs (local Qwen 7B on RTX 4090)
What needs work ~50% false positive rate – AI over‑optimizes for "pain" and under‑optimizes for "sellable"
No feedback loop yet – the system doesn't know which "BUILD" posts actually convert
Occasional hallucination (e.g., "pet programmer" suggestion from unrelated post)
Log entries File Stage One sentence How-to-Find-Real-Problems-Worth-Solving-on-Reddit.md Discovery & evaluation Building the original two-stage filter concept on Reddit (now superseded by HN) What-a-9k-Founder-Taught-Me-About-Demand-Validation.md Evaluation Adding cross-platform signals (G2, Upwork) to strengthen BUILD decisions instant-pricing-intake.md Action Designing a pricing format for manual demand analysis reports 2026-07-01-local-ai-demand-pipeline.md Pivot & build Switched from Reddit to HN, integrated Qwen 7B, achieved working pipeline with structured pain point extraction What this is A raw, unpolished record of a solo builder figuring things out in public
Some files are newly written; others are older work cleaned up and archived here
No products, no landing pages, no waitlists — just an engineering logbook
What's next Run 30 posts manually label every output (build a truth dataset)
Add "solutionizability" dimension to Stage 2 prompt
Build feedback_loop.py to auto‑tune scoring weights based on actual conversion data
Product generator: accept a BUILD post → output full Notion template ready for Gumroad
Last updated: 2026-07-01