fastpaca What happens when LLMs hit production?
I write about memory systems, context management, agents, evals, and the parts of AI infrastructure that break when real users show up.
I'm Seb. I've spent years around trust & safety, platform work, and AI tooling; fastpaca is where I write down what keeps being useful.
Writing
All posts →Building on a Moving Train
AI infrastructure feels brittle because every primitive beneath it moves. Here's what I learned building and killing infrastructure in a field that reinvents itself every quarter.
Let's Build an AI Assistant That Remembers
A practical walkthrough of building an assistant with persistent memory using Cria.
Ultimate Guide to LLM Memory
How do you add memory to your agent or LLM? What works and what does not? How do you use multiple memory systems at once to cover each others weaknesses?
Design Your LLM Memory Around How It Fails
Not all context is sacred. Design your agent's memory around what happens when critical information gets dropped.
Things I Built
GitHub →cria
Prompt architecture for model stacks that refuse to stay still.
Mostly me trying to keep prompts legible while models, providers, and memory layers change underneath.
pacabench
Agent benchmarking without hiding the state.
Local-first eval runs with isolated execution, reproducible inputs, and enough state left behind to inspect what actually happened.