I build production AI systems — voice platforms, LLM infrastructure, and the quiet, durable backends that make them actually work in the wild.
As Founding Engineer at Better Collision Centers, I'm shipping an omnichannel AI assistant on Twilio SIP that handles real customer conversations across voice and SMS — with the boring, important things (retries, idempotency, monitoring) treated as first-class concerns.
The interesting work sits underneath: an LLM-driven scheduler with feedback loops that has pulled cycle time from 15 days to 10, and a backend where model outputs cleanly trigger downstream operational decisions. AWS, RBAC, audit-ready.
Quantum-native AI cognition. An experiment in whether a model can think — not just respond — through a qubit substrate.
The setup: an 8-qubit register runs alongside the prompt. As the model reasons, the circuit ticks. I capture those ticks as a kind of cognitive signature, then replay them against an unrelated question to see whether a thought-pattern — a mood — can be carried across prompts.
Open questions: can the quantum state genuinely steer the model's conclusions, or am I just measuring noise dressed up as cognition? Early days, mostly notebooks, careful controls, and a healthy skepticism.