Technical
notes.
Writing about the field, not about us. Every technical claim carries a clickable source with the date it was accessed; every note argues a thesis rather than surveying. A note without footnotes does not go on the site.
A source, or silence.
The field moves faster than anyone’s memory, so nothing here is written from recall. Anything presented as current was checked on the day of writing — with a link and an access date on every entry.
A thesis, not a survey.
A survey without an opinion is generated content. Every note defends a single sentence and says outright where certainty ends — a caveat written into the text is stronger than a smooth sentence with nothing behind it.
What is not here.
Nothing about salaried work or projects covered by someone else’s IP. Nothing that cannot be checked through a link. No keyword-driven content — this is not a blog chasing traffic, it is an engineer’s registry of thinking.
Topic queue.
pool approved · ordered by thesis approvalRetrieval
- modular RAG vs monolith
- agentic RAG: when it pays off
- GraphRAG: structure vs over-engineering
- BM25 + vector hybrid, reranking
- long context vs RAG
Agents
- MCP as a tool standard
- memory as architecture
- context engineering
- orchestration and the human entry point
Documents
- OCR + LLM: the validation loop
- enforcing response structure
- when NOT to use an LLM
Production
- evaluation: gold sets, LLM-as-judge
- observability, A/B, rate limits
- model routing and semantic cache
- local vs API: GDPR
- extreme quantization
pending approvalLong context did not kill RAGPasting everything in costs you three times over: price, attention decay in the middle of the context, and time to first token.002thesis
pending approvalIn extraction, the LLM should read — not calculate, not decideSchema enforced at decoding, deterministic validation, a human approves only the exceptions.003thesis
pending approvalMost „AI in the company” projects should be ordinary codeIf the rules can be written down, an LLM adds cost, latency and nondeterminism to a problem solved forty years ago.
How a note reads.
A different material from the model cards and results tables: serif in a single column, footnotes instead of threads, caveats in boxes, sources with access dates at the end. Three elements you will see in every entry:
footnote → source with a date
A technical claim ends in a reference[1], and every entry in the source list looks like this:
[1] author/title · link· accessed: day of writing
thesis up front
Every entry opens with a thesis in an accented box — one sentence the text defends. As in a paper: the claim first, then the proof.
caveat in the open
Where certainty ends, the text says so openly — the dark box is part of the format, not fine print. The rule was tested on the 76.50% figure in the paper.