Halide Yılmaz Production
5 posts
Production8 minAI Incident Reporting and Post-Market Monitoring: The 15-Day, 10-Day and 2-Day Clocks
Serious incident reporting runs on a 15-day clock, 10 days in the event of death and 2 days for a widespread infringement. What logging infrastructure, severity taxonomy and on-call path an ML team needs in order to meet those clocks at all.
Production10 minEmbedding Drift: Monitoring for Drift in the Foundation-Model Era
Why a per-dimension KS test fails on a 768-dimensional embedding, what to measure instead, how an embedding model version bump silently invalidates your reference window, and why drift is not degradation.
Production10 minKnowing Your Model Broke Before the Labels Arrive: Performance Estimation Without Ground Truth
If the labels arrive in ninety days, is your model broken now? How confidence-based performance estimation works, the assumptions it rests on, and why it goes quiet in exactly the scenario people reach for it.
Production10 minFeature Stores and Training-Serving Skew: Why Your Model Behaves Differently in Production
Offline the model scores well, in production it behaves differently. The four concrete sources of training-serving skew, why point-in-time correctness is really leakage prevention, and when a feature store earns its operational cost.
Production8 minModel Monitoring and Drift in Production: Concept Drift, Performance Decay, and Retraining Triggers
A practical guide to data and concept drift in production ML models: detecting performance decay, retraining triggers, and keeping your production ML healthy.