2 min read
RAG is a retrieval problem, not a prompting problem
Two-thirds of the accuracy gain in our production RAG system came from changes upstream of the model. Here is the measured breakdown, and what to fix in what order.
The blog
Writing on applied AI.
Architecture notes, evaluation methodology and post-mortems from building language and vision systems that run in production.
1 min read
How to evaluate an LLM system without fooling yourself
Vibes-based evaluation is how teams ship regressions. A practical framework: golden sets, calibrated judges, regression gates, and the failure modes of each.
1 min read
Fine-tuning vs RAG: a decision procedure, not a debate
The question is not which is better. It is which failure you are trying to fix — missing knowledge, wrong behaviour, or unacceptable cost.