About ALLM Academy
A free, open course on production-LLM engineering — the systems work around the model: harness design, context windows, structured output, tools, MCP, state, memory, orchestration, semantic and graph-backed RAG, inference, evals, observability, cost, and safety. 15 modules across 4learning paths and 32 focused topics, taught with accessible 2D diagrams, worked examples, quizzes, and optional audio.
How the daily briefing works
This is a fully static site, so there is no server fetching news at runtime. Instead a scheduled CI job runs a pipeline that fetches from public sources (Hacker News, arXiv, GitHub Search, and several blogs via RSS),dedupes / filters / ranks by relevance and recency, then curates the top items into short teachable lessons mapped back to the curriculum. With an LLM key it writes a polished summary + a micro-quiz; without one it falls back to a deterministic templated summary (flagged auto-summary). The result is committed and the site redeploys.
GitHub discovery is keyless and bounded to one result from each of four cached searches for recently pushed GraphRAG, agent-memory, harness, and PrimeIntellect-ai/prime-agent repositories. Current stars and forks help rank discoveries; they are not treated as evidence that a project is novel, correct, or effective. Every lesson retains its direct source URL, source type, and available publication or repository-activity time.
Honest limits
- Numbers, prices, and benchmarks in lessons are illustrative ranges — verify against primary sources.
- Some feeds (e.g. Reddit) block automated requests and may be skipped; the pipeline degrades gracefully.
- X/Twitter has no stable keyless feed; it is intentionally omitted (optional enrichment is a stretch goal).
- "Quiet day" cards appear when nothing clears the relevance bar — that is by design, not a bug.
Your data
Progress (completed lessons, quiz scores, streak) lives only in your browser's localStorage. No account, no tracking. Export it for safekeeping or move it between browsers.