Technical writing on systems I have built. Infrastructure, ML pipelines, and the engineering decisions behind them.
Why single-document inference wastes 85% of GPU capacity, and how type-based grouping, page-count batch sizing, and dual-path routing brought utilization from 15% to 70%+ with zero OOM incidents.
How we built compliant-by-default infrastructure at Archv. Column-level encryption, row-level access control, immutable audit logs, and zero-trust service mesh. All automated, no manual checklists.
Aggregate accuracy dashboards hide localized failures. We built alerting on data drift, prediction confidence drops, and silent model degradation. Alerts fire before users report problems.