markdown · thesis
Build AI systems that survive production.
A running notebook on RAG, tool-using agents, and honest evals — what worked, what broke, and the numbers behind both. Every post below is a cell, executed in order, newest at the top.
runtime 3 cells 0 still running 3 namespaces
- Evals active cell
Evaluating LLM outputs without fooling yourself
Vibes are not a metric. A practical framework for measuring whether your LLM feature is good — grounded in test sets, rubrics, and the honest limits of LLM-as-judge.
- Systems
MCP: giving language models real tools
The Model Context Protocol turns an LLM from a text generator into something that can read your files, call your APIs, and act. Here is what it is and how to build a server that does not misbehave.
- Evals
RAG pipelines that actually work in production
Most RAG demos fall apart on real corpora. Here is the retrieval stack that survives messy documents, ambiguous queries, and users who paste in nonsense.
run all 3 cells in the archive →