A learning-in-public log tracking current AI engineering — the models, tools, protocols, and harnesses that actually shipped in 2025–2026, not last year's hype.
I'm a Senior AI Engineer working on agentic infrastructure and AI governance. This repo is where I take notes, write guides, and keep my own map of the fast-moving AI stack up to date. Everything here is written to be useful to someone else picking up the same tools.
The AI space moves faster than any single blog post can keep up with. Frameworks that were the default six months ago are already legacy. This repo is my attempt to stay current by writing things down — the classic "learning in public" loop:
- Read / try a new tool, model, or protocol.
- Write a short note (
til/) or a full guide (guides/) explaining it plainly. - Revisit and correct earlier notes as things change.
If a guide here is wrong or out of date, open an issue — that's the whole point.
This isn't only agent frameworks. It tracks the full current stack:
- Agents & orchestration — Claude Agent SDK, LangGraph, CrewAI, multica, openworker
- Protocols — MCP, Agent Skills, agent-to-agent
- RAG & retrieval — LlamaIndex, Haystack, RAGAS, vector DBs (Milvus / Qdrant / Weaviate)
- Inference & serving — vLLM, Groq, Cerebras, on-device quantization
- Fine-tuning — Unsloth, LlamaFactory
- Evals & observability — DeepEval, RAGAS, tracing
- Harness engineering — the durable patterns behind all of it
See ROADMAP.md for the full backlog.
In-depth, plain-language explanations of current tech.
| Guide | Topic | Status |
|---|---|---|
| MCP explained | Model Context Protocol — the USB-C of agent tools | ✅ |
| Agent harness anatomy | The control-plane loop around a model | 🔜 |
| vLLM in practice | High-throughput LLM serving | 🔜 |
| RAG eval with RAGAS | Measuring retrieval quality | 🔜 |
| Unsloth fine-tuning | Low-VRAM domain adaptation | 🔜 |
Short daily notes live in til/, one folder per day (til/<date>/NN-slug/).
One small, real thing learned per entry — no filler.
This repo is fed by an openly automated daily routine — I'll say that plainly
rather than pretend each post is hand-typed at dawn. A scheduled cloud agent runs
every morning, reads a fresh digest of real current sources (scripts/digest.py),
and writes across three streams, each grounded in and citing today's real sources:
posts/<date>/NN-slug/— full bilingual (RU + EN) explainers of the day's most notable items, illustrated with the source's own image (else a generated diagram);til/<date>/NN-slug/— short "today I learned" notes on smaller items;guides/<date>/NN-slug/— deeper plain-language explainers of one concept.
Several folders per stream per day. The content is real and grounded in today's sources; the automation is the point, not a disguise.
Where the digest comes from. The routine's cloud sandbox blocks outbound traffic to
almost everything (arXiv, vendor blogs, Hacker News). What it can reach is
raw.githubusercontent.com, so feeds.txt leads with community RSS mirrors hosted
there — HF trending papers (with abstracts), GitHub trending/ranking repos, HF blog —
that are rebuilt daily. digest.py fetches those in-sandbox; the arXiv API and vendor
feeds below them enrich the digest only when the script runs on an open network. No
GitHub Action and no API key are involved — the routine is fully self-contained.
Evergreen deep-dives stay at the top level of guides/ (e.g.
MCP explained); the dated daily explainers land in
guides/<date>/ alongside them.
The digest fetcher is the only moving part you'd run by hand; it needs no API key:
python3 scripts/digest.py # fetch today's digest (arXiv + feeds.txt + trending repos)MIT — take anything useful.