PT-Edge is a precomputed inference cache designed to assist AI infrastructure decisions. It tracks more than 220,000 open-source AI projects across GitHub, PyPI, npm, Docker Hub, Hugging Face, and Hacker News, scores their quality daily, and publishes the results as a catalog site containing more than 220,000 pages.
Purpose: The site serves two audiences: first, AI agents reading pages on behalf of humans, where the page content is structured, front-loads information, and is machine-readable, so agents can obtain confident, citable recommendations in a single visit; second, humans reading directly, where pages are navigable, trustworthy, and contain original analysis.
Key capabilities:
- Data ingestion: a daily pipeline pulls GitHub statistics, package download counts, releases, HN posts, Hugging Face models/datasets, and registry data.
- Scoring: computes a composite quality score (0-100) across four dimensions: maintenance, adoption, maturity, and community. Quality tiers include: Verified (70-100), Mature (50-69), Emerging (30-49), Experimental (10-29).
- Enrichment: LLM-generated technical summaries, practitioner-oriented assessments, and comparative analysis based on READMEs.
- Publishing: generates a static site across 17 domains, including structured data, internal links, and freshness signals. The catalog site is located at mcp.phasetransitions.ai, covering 17 domains and more than 2,400 categories, updated daily.
- Observation: analyzes bot traffic to form a "Demand Radar."
Usage and technical requirements: The system uses Python, FastAPI, PostgreSQL + pgvector, LLM enrichment (supporting multiple providers), generates a static site via Jinja2 + Tailwind CSS, and is hosted on Render. The entire system runs on a single server instance, with a monthly cost of less than $300. It provides MCP tools and a REST API for programmatic access. This is a production system with no staging environment; the database is a PostgreSQL instance larger than 1GB, and queries hit real data. Development setup and safety rules are in docs/development.md.
Demand Radar: Every bot visit on the site is potential intelligence. Access logs carry three layers of signals: indexing bots (Meta, Anthropic, Amazon, Google, Perplexity, OpenAI) reflect what each AI company believes will be valuable in future model weights, and each bot has a unique crawling strategy that can reveal the parent company's priorities; user-behavior bots (ChatGPT-User, OAI-SearchBot, Perplexity-User) reflect what real humans are currently asking AI, and each hit represents a practitioner making technical decisions through an AI intermediary; human visitors reflect what people find directly through search engines. Demand Radar extracts these signals and feeds them into content prioritization, ultimately to be implemented through a trained ML model rather than hand-tuned weights. The full analysis is in scratch/demand-radar/.
Documentation and license: 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. The project was built by Graham Rowe and is licensed under MIT.
