How I stopped burning money on LLM calls with Promptfoo
I used Promptfoo evals to tune cheaper models, catch regressions early, and stop wasting money on flaky LLM calls.
What I learned about prompt injection defense in production
Lessons from shipping multiple defense layers in an AI fitness app, and the tradeoffs behind each one.
Building a visual debugger for agent tool chains
A TraceCollector that records every event in the agent's lifecycle and feeds them to an interactive flow graph.
Taming a monolithic system prompt
A composable prompt architecture where named segments get assembled at request time based on context.
Letting an LLM loose on my database
A structured query DSL that lets the model express intent as JSON while the server compiles it into safe, parameterised SQL.
Self-healing tool calls with a cheap LLM
Intercepting validation errors and asking a cheap model to fix the JSON instead of failing the whole call.
Cutting my agent's response time in half
I halved my AI agent's response time by not loading tools it doesn't need.