Full-Stack Engineer who builds AI-assisted workflows.

The hard part isn't extraction. It's deciding how much automation each email earns: classify intent, apply scoped account memory, validate against the reservation schema, surface missing fields instead of guessing, then either auto-import complete requests or route incomplete ones to an operator. When an operator fills missing details and the booking is created, a learn agent saves verified addresses and preferences for future requests.

A support reply is a diagnosis problem, not a generation one. Two things have to come together before anything is written. First the system assembles context: MCP tools pull account + reservation state while an OpenAI Agents SDK agent retrieves internal wiki knowledge, and the two drive issue classification. In parallel it searches resolved emails by embedding, refining the query until matches are genuinely relevant — but those precedents only point toward a fix, they aren't the fix. A diagnose step turns precedent plus account context into the real cause (e.g. an account number that was never filled in) and the concrete action the user must take in GNet or the reservation system. The grounded draft is the product of both lanes — assembled context and the diagnosis — which the human reviews and sends. A diff against the draft then feeds whatever the human changed back into both stores: new precedent into resolved emails, new how-to into the wiki.

Steering is a pipeline, not a trick. Train the GPT from scratch, train sparse autoencoders on six of its layers, and read the geometry: every concept is superposed, so no single clean feature can carry an override. Instead, two cross-checked signals find the profession features, an adversarial skeptic panel throws out the morphological and prediction artifacts, and the survivors become a curated set with recommended strengths. A formula — layers × features × signed strength — is injected into the residual stream via forward hooks at generation time, and the answer changes with zero weight edits.