No matches. Try a different word.
5 Oct 2026
I compared three ways of using Jev to check cached answers, then tested it against GPT-4.1 mini. The checks were cheaper and faster, with a trade-off in how many answers could be reused.
Read in the notebook →13 Sept 2026
A step-by-step journey from vectors and the law of cosines to why the dot product of normalised vectors measures directional similarity.
Read in the notebook →28 Aug 2026
A first-principles look at how GPT-style language models turn raw text into token IDs, token embeddings, and positional embeddings.
Read in the notebook →13 Aug 2026
The PEAS framework offers a way to design AI agents from the outside in by defining success, environment, actions, and observations before writing code.
Read in the notebook →30 Jul 2026
Most research on agent memory focuses on remembering the user. I built a browser agent that learned from its own runs, and the lessons that transferred surprised me.
Read in the notebook →10 Jul 2026
Stronger models help, but agent reliability may depend just as much on better context systems around them.
Read in the notebook →6 Jul 2026
Production-like evals revealed the retrieval architecture I actually needed and reminded me that confidence in agentic systems has to be earned, not borrowed.
Read in the notebook →19 Jun 2026
When agents can read private data, process untrusted content, and communicate outward, prompt injection becomes a much more serious security problem.
Read in the notebook →17 Jun 2026
Plan caching reuses planning templates across similar agent tasks, cutting cost and latency without throwing away accuracy.
Read in the notebook →5 Jun 2026
Bigger context windows do not remove failure modes. They create new ones when we stop being intentional about what goes into an agent's context.
Read in the notebook →21 May 2026
I benchmarked five embedding models across four NanoBEIR datasets and found that bigger embeddings did not always produce better retrieval.
Read in the notebook →19 May 2026
Bi-encoders make retrieval fast, but cross-encoders expose why reranking matters when meaning depends on the query.
Read in the notebook →11 May 2026
GGUF made local LLM inference feel practical by packaging model weights, vocabulary, hyperparameters, and architecture metadata into one runnable format.
Read in the notebook →27 Mar 2026
Repeated user intents can quietly inflate LLM cost and latency. Semantic caching helps, but production use comes with trade-offs.
Read in the notebook →