Bug: sort=stars returned the same viral repos every cycle (ponytail 77K⭐
dominated for 26 days). A 5-day-old repo with 3K stars (600/d) was
ranked below it despite having nearly 2x the growth velocity.
Fix:
- Fetch repos created in last 30 days (was 14)
- Re-sort by velocity (stars/day) instead of total stars
- Score formula: 70% velocity + 30% absolute stars
- Added 3rd query for breadth (autonomous agent OR AI tool)
- Added stars_per_day to raw_metadata for queryability
- Added velocity field to metadata
- adapters/huggingface.py: HF API adapter (models + datasets)
- Dual fetch: sort=likes (popularity) + sort=lastModified (fresh)
- AI relevance filter: pipeline_tag, library_name, tag matching
- Score: adoption (likes/downloads log-scale) + relevance (pipeline/library/tags)
- score_type: actual (real likes/downloads from HF API)
- Cross-source signal: GLM-5.2 top on both HN and HF
- Wired into ENABLED_SOURCES + verification in pipeline.py
- Live verify: DB likes match live API exactly (GLM-5.2: 3607, DeepSeek-R1: 13448)
- 99 total entries across 5 sources, all pipeline green
- adapters/hackernews.py: HN Firebase API adapter with AI keyword filtering
- Word-boundary matching to avoid substring traps (Britain/Guinea)
- Score: log(points) + log(comments), actual HN scores
- Wired into ENABLED_SOURCES + verification in pipeline.py
- Live test: 10 AI/ML stories fetched, all clean
Epictetus