§ 04
Writing
- 2026-07-05Accurate Offline, Useless In-App: The Tradeoff Nobody Grades You OnDuring a volunteer internship at Bluesense AI I integrated Hugging Face face-analysis models into a real product and built HRV/ECG signal pipelines. The lesson that stuck wasn't about accuracy — it was that a model's benchmark score and its in-app usefulness are two different numbers, and only one of them ships.
- 2026-07-04Scheduling a Retail Floor with CP-SAT: An Honest Shift-Optimization ArchitectureHow I modeled hourly staff placement in a ZARA store as a constraint-programming problem with Google OR-Tools CP-SAT, served it over FastAPI, validated it against ten periods of real human decisions — and how the solver then outgrew the hackathon and became the scheduler my section actually uses every day.
- 2026-07-04When Your Model Scores a Perfect 1.0, Don't Celebrate — InvestigateMy first conversion model scored a perfect ROC-AUC of 1.0. It wasn't skill — it was target leakage. This is how a six-scenario ablation found the culprit, and why the honest 0.71 turned out to be worth more than the fake 1.0.
- 2026-07-04What Coordinating Shifts in Orquest Taught Me About ConstraintsI'm a part-time Sales Advisor who ended up coordinating shift planning for a large ZARA store through Orquest. These are the scheduling lessons from the inside — contract guarantees, fixed budgets, and why 'no changes after publish' is a feature, not a bug.