grahamrowe82/pt-edge

PT-Edge is a precomputed reasoning cache for AI infrastructure decisions. It tracks 220,000+ AI repos across GitHub, PyPI, npm, Docker Hub, HuggingFace, and Hacker News, scores them daily on quality, and publishes the results as a 220,000+ page directory site.

The site serves two audiences: AI agents reading pages on behalf of humans (structured, front-loaded, machine-readable) and humans reading directly (navigable, trustworthy, original analysis). Every page is designed so an AI agent can land on it and walk away with a confident, citable recommendation in one pass.

Every major AI lab's crawl infrastructure treats the site as a primary data source. The access logs are themselves an intelligence layer — see Demand Radar below.

Directory site: mcp.phasetransitions.ai — 220,000+ pages across 17 domains with 2,400+ categories, updated daily.

Built by Graham Rowe

How It Works

1. Ingest — daily pipeline pulls GitHub stats, package downloads, releases, HN posts, HuggingFace models/datasets, and registry data
2. Score — composite quality score (0-100) from four dimensions: maintenance, adoption, maturity, community
3. Enrich — LLM-generated technical summaries, practitioner-focused assessments, and comparison analyses from READMEs
4. Publish — static site generation across 17 domains with structured data, internal linking, and freshness signals
5. Observe — bot traffic analysis reveals what the AI ecosystem values (Demand Radar)

The entire system runs on a single server instance for under $300/month.

Quality Scoring

Dimension	Max	Signals
Maintenance	25	Commit activity (30d), push recency
Adoption	25	Stars (log scale), monthly downloads, reverse dependents
Maturity	25	License, PyPI/npm packaging, repo age
Community	25	Forks (log scale), fork-to-star ratio

Tiers: Verified (70-100), Established (50-69), Emerging (30-49), Experimental (10-29)

Demand Radar

Every bot hit on the site is latent intelligence. The access logs carry three layers of signal:

• Indexing bots (Meta, Anthropic, Amazon, Google, Perplexity, OpenAI) — what AI companies think will be valuable in future model weights. Each bot has a distinct crawl strategy that reveals its parent company's priorities.
• User-action bots (ChatGPT-User, OAI-SearchBot, Perplexity-User) — what real humans are asking AI right now. Each hit represents a practitioner making a technology decision through an AI intermediary.
• Human visitors — what people find through search engines directly.

The Demand Radar extracts these signals and feeds them into content prioritisation — eventually via trained ML models rather than hand-tuned weights. See scratch/demand-radar/ for the full analysis.

Stack

Python, FastAPI, PostgreSQL + pgvector, LLM enrichment (multiple providers), static site generation via Jinja2 + Tailwind CSS. Hosted on Render. MCP tools and REST API for programmatic access.

Development

This is a production system with no staging environment. The database is a live 1GB+ PostgreSQL instance — queries hit real data. See docs/development.md for setup notes and safety rules.

Documentation

• docs/strategy.md — strategic positioning and reasoning
• docs/roadmap.md — phased build plan
• docs/design/worker-architecture.md — task queue and worker design
• docs/development.md — development setup and database safety
• scratch/demand-radar/ — access log intelligence analysis

License

MIT — see LICENSE.