Peter J. Wyatt
I run restaurant operations and I build the software that runs them, and I hold both halves to the same standard: a number ships only if it survives being attacked.
Five years in a multi-unit quick-service restaurant group taught me where operational systems break, because I was the person they broke on. The work below is what I built from the other side of that, and how I validated it.
Forecasting restaurant demand, and proving the number honestly
A demand forecasting engine that hits 92.1 percent daily accuracy, validated by rolling-origin backtesting across 44 weeks of real hourly sales rather than by a shuffled cross-validation that would have leaked the future into the training set. Covers how it is scored against a seasonal-naive baseline, why the hourly number is lower than the daily one and why a higher hourly number is not physically available, and where forecasts genuinely fail: holidays, local events, promotions, and a weather signal I tested and could not make help.
Read the forecasting case study
The bug that was lying for me
An AI form filler I built answered "Yes" to any question matching "experience with." It was a plausible heuristic that fabricated qualifications, and it had already run across 215 automated submissions. This is the audit that found it, the two other defects in the same pass, and the guardrails that replaced them, alongside the rest of the infrastructure: eighteen ingest adapters over a 73,431-posting pipeline, a 4,194-line agentic assistant with nine tools and RAM-aware switching across 3B, 7B and 14B models, Model Context Protocol servers, and monitoring that runs on an 8 GB laptop. Plus the LLM failure modes I have actually hit, and what each one required in code rather than in a prompt.
Read the agentic infrastructure case study
About
Operations leader and builder, with a standing interest in what happens when a system is allowed to speak on your behalf.