Engineering notes
on production AI.
Long-form essays on building production AI systems — agentic orchestration, tool reliability, evaluation, and the infrastructure that keeps it running at scale.
Can We Really Make AI Safe?
Can we really make AI safe when we do not fully understand the model, and we are steadily giving it more access to the world?
Fireworks AI ZDR: How to Set It Up Without Missing the Exceptions
Fireworks has a good zero-retention default for open-model inference. This guide covers the API settings, cache isolation, residency, and feature boundaries that decide whether that is true for your application.
Where Should the AI Actually Go?
A model can be capable of doing a task and still be placed in the wrong part of the system. I learned this while building AI for a legal workflow.
The Ship You Can't Dock: Architectural Debt in the AI Era
In the fast-moving AI space, architectural debt isn't just about cutting corners—it's about reasonable decisions being invalidated by a shifting environment.
Scaling to 1,500 Concurrent Users: PgBouncer and Null Pooling
A deep dive into why application-level pooling fails for long-running AI workflows and how to implement PgBouncer with statement-level pooling to handle 30x the load with 10x fewer resources.
Zero Data Retention (ZDR) for LLM Providers
A practical guide to keeping your data private when using LLM APIs. Covers zero-retention endpoints, self-hosting, compliance requirements, and data protection patterns.
What Gemma 4 Actually Does Differently
Gemma 4's 31B model is outscoring systems with 10x more parameters on Arena Elo. Here's the architectural reasoning behind why that's possible.
Injection Is Not Influence: The Illusion of LLM Memory
After three years of building LLM applications, I've learned that LLM memory is fundamentally different from human memory.