Writing
Short posts. Plain words.
What I’ve learned about testing, APIs, databases and LLM features. Each post starts with the short version.
8 posts
- 2 min
LangGraph and Langfuse, in plain words
My notes while learning two tools for AI apps. LangGraph decides what the app does next. Langfuse records what it actually did. You need both.
- 3 min
How to build an API for 100,000 requests a second
The number alone means little. What matters is the kind of traffic, how many requests are in flight at once, and what your system does when it is over its limit.
- 3 min
Your API is slow because of the database
A slow API is rarely slow because of app code. It is usually a missing index, a query inside a loop, a pool that is too big, or paging that gets slower the deeper you go.
- 2 min
Why blocking too much breaks an AI guardrail
A guardrail can fail two ways: it lets an attack in, or it blocks a normal question. The second one is quiet — and it is the one that gets guardrails switched off.
- 2 min
How to test an LLM feature in CI
LLM answers change wording on every run, so normal tests break. Check structure, facts and meaning in layers, and pass on a percentage — that gives you a real gate.
- 2 min
API tests that don’t break when the API changes
One test class per API breaks every time a field is renamed. Describe each test as data instead, and a rename becomes a one-line config change.
- 2 min
When RAG gives a bad answer, check the search first
When a RAG assistant answers badly, people blame the model and rewrite the prompt. Usually the right document never reached the model at all.
- 2 min
How to fix a slow, flaky test suite
A suite that people rerun until it turns green is no longer a test suite. Flaky tests come from about five causes, and each one has a clear fix.