Making the Foundation Model Ecosystem Legible
The modern AI stack moves quickly across two worlds that rarely share a single map: LLMs & agent systems (KV caches, MoE, vLLM, MCP, LangGraph) and image & video generation (DiT, VAEs, checkpoints, Safetensors, LoRAs, ControlNet, ComfyUI). FoundationIndex unifies both into a chronological, source-grounded, cross-linked field guide inspired by Ekşi Sözlük's rabbit-hole hyperlinking and Hacker News's recency.
System Architecture & Roadmap
Built with TanStack Start, TanStack Router, TanStack Query, and Tailwind CSS, statically prerendered and served on Cloudflare Workers. All subjects live in Git-backed YAML validated at build time with Zod—enforcing zero broken [[cross-links]] and automatically generating bidirectional backlinks and the d3-force knowledge graph.
An automated daily Python workflow using LangGraph and Pydantic scans Hacker News, Hugging Face Daily Papers, arXiv, and major lab feeds—matching items against existing subjects, detecting missing concepts (gaps), verifying citation grounding, and opening a structured GitHub PR for human review with an automated pytest eval suite.
A Remotion (React/TypeScript) pipeline turning structured subject data and graph neighborhoods directly into 60-second YouTube Shorts and weekly long-form landscape explainers—embedded back onto each subject page.
Syncing the Git-backed knowledge base into PostgreSQL with pgvector behind a FastAPI service for hybrid semantic search, citation-grounded “Ask the Landscape” Q&A via TanStack Query, and interactive VRAM / KV-cache calculators.
Built by Üstün Özgür
Principal Software Engineer, Full-Stack Product & Applied AI Engineer (15+ years with US startups across founding engineer, CTO, VP of Engineering, and Principal/Lead roles; M.Sc. in Electrical & Electronics Engineering, Bilkent University). Currently exploring AI/ML Engineering and Senior/Staff Backend roles.