Silicon River

Model releases move far faster than anyone can track by hand. Every day brings new weights, new benchmark scores, new pricing — scattered across Hugging Face, OpenRouter, vendor blogs and X. Answering “what actually happened this week” is unreasonably expensive.

Silicon River is my answer to that. Not another static list, but a 3D spiral timeline: nodes can be clicked and locked for focus, showing full model metadata and pricing, with multi-dimensional filtering by time range, provider, license and keyword.

How the data gets in

• Python scrapers pull the public Hugging Face and OpenRouter catalogues on a schedule; every sync writes a sync log, so it stays auditable and traceable
• Structured results land in PostgreSQL, with caching and incremental updates
• FastAPI exposes /api/timeline and friends to the frontend

What “AI-enhanced” actually adds

This is the layer I put the most effort into, because it decides whether the timeline is data shuffling or genuinely useful:

1. Real-time web search (Brave Search API) — pulls public information around a release, solving the “the model is on the leaderboard but nobody wrote down what it is” problem
2. Multi-leaderboard cross-validation — queries SWE-bench, LMSYS Arena and the Hugging Face Leaderboard together, so no single leaderboard’s bias steers the result
3. LLM profiling — DeepSeek generates model descriptions from the web context plus leaderboard data
4. Four-tier milestone detection — strong keywords → weak keywords → timeline rank → leaderboard rank, narrowing layer by layer to surface the releases that actually mattered

Engineering trade-offs

• Degrade first: when any external dependency (search, leaderboards, LLM) is down, the service returns a degraded result instead of erroring — the timeline itself is always available
• Test coverage: 75+ test cases across scrapers, leaderboard clients, web search and end-to-end enhanced analysis
• Delivery: one docker-compose brings it up; the frontend ships as its own nginx image, plus a frontend-only compose for UI work
• Observability: phase four added production-grade performance monitoring and operational runbooks

Status

Actively iterating. Backend is Python + FastAPI + PostgreSQL; frontend is React + Three.js.