Blog
Deep dives into the engineering tradeoffs behind the systems I ship.
How I designed a stateful LLM workflow around durable checkpoints, safe replay semantics, bounded review loops, and a storage path for state that outgrows a database record.
AI doesn't tell you when it's guessing. It generates text that sounds right instead of text that is right. What hallucinations look like, why they happen, what goes wrong when you trust them in high stakes work, and how to engineer around them.
The Calc 2 wall is where most tech kids quit. The ones who push through and finish a math degree walk out with the most flexible, future proof ticket into software, data, and AI.
LLMs hallucinate. In a compliance pipeline, one hallucinated AWS service name can taint an entire audit trail. Here's how I stopped trusting the model as an oracle and rebuilt it as a scoped classifier wrapped in deterministic code that doesn't negotiate.