Notes on applied AI, software engineering, and technology.
How much more the same text costs in Portuguese than in English, measured on 31 articles of this site with nine tokenizers, and what to do about it in prompts, budgets and RAG.
14 min read
Every way to run a Claude agent while nobody is watching, from claude --bg to Managed Agents, and a tested nightly setup with permissions, hooks, budgets and a report you can trust.
27 min read
PageIndex swaps the vector database for a document tree and an LLM agent that navigates it. I benchmarked it against two vector RAG pipelines on FinanceBench: better answers, 40 times the cost.
22 min read
I built a working phishing screener on Laya, the open-weights decision model: five typed questions in one forward pass, deterministic checks on the headers, and a policy that needs evidence before it quarantines.
8 min read
How I built an event-driven AWS pipeline that transcribes, indexes, and queries long-form video, and why the model never gets to write its own citations.
8 min read
How I built a RAG MVP with FastAPI, Celery, Qdrant, Azure AI Foundry, and streaming, and what production still requires.
7 min read
Instrument an agent loop with the OpenTelemetry GenAI conventions, redact content before export, and turn token counts into cost per feature and tenant.
36 min read
How contrastive training shapes an embedding model, and the prefixes, pooling, normalization and truncation details that decide whether it retrieves well on your data.
29 min read
Decide with evals when fine-tuning is worth it, then build LoRA from scratch in PyTorch and train a 135M model to triage support tickets.
33 min read
What quantization does to an LLM's weights, activations and KV cache, and how to measure what you lost before you ship it.
37 min read