6 posts · forward deployed engineering, production AI, inference, evals
What it actually looks like to wire Azure SQL, Azure ML, Cosmos DB, Azure Functions, and a Static Web App into one working system. Including the part where 82% accuracy means the model does nothing.
Dhaga & Co. had 44% of returns landing in a free-text field their category head could not scale through manually. Three agentic AI patterns turned that into actionable root cause intelligence for $2.50 a week.
A payroll SaaS company asked for a chatbot to cut headcount by 30%. The data told a different story. A walk through what problem discovery actually looks like in the field.
The prototype-to-production gap is not a model problem — it is a systems problem. Here are the four layers where AI deployments break after the demo looks great, and what to do about each one.
FDE interviews are deliberately open-ended. Most engineers freeze because they are waiting for a spec that never comes. Here is the framework that removes the freeze — a repeatable system for reasoning through ambiguous problems out loud.
The bottleneck in AI has moved from building to deploying. The engineers who win in this era are not the ones who write the best code — they are the ones who close the gap between a working model and a working system in a customer's hands.