Evidence-grade assurance for AI in production.

NVIDIA GenAI Credentialed (NCA-GENL) · MS Information Systems, Northeastern University (2024) · NSRI / underwater autonomy defense research · combined experience across data engineering, distributed pipelines, and production AI.

My work sits at the intersection of information systems analysis and production AI assurance. It starts with the same question every serious AI team eventually faces: can you prove this system is reliable, affordable, safe, observable, and defensible? That is not a prompt problem. It is an evaluation-harness, routing, and control-matrix problem.

KYC Copilot, my own product—built commercially, then open-sourced—is the clearest public example. It is not a client engagement; the benchmark comes from the project's test dossier, not a client production record. The compliance workflow runs at 93% lower manual review time (14 minutes automated dossier plus HITL versus 3.5 hours manual), with 159/159 unit tests, 24/24 evaluation scenarios, F1 1.000, KYT macro-F1 1.000, and a 0.800 low-cost routing ratio: 80% of low-risk cases route to tier t2, while high-risk PEP/sanctions cases route to tier t4 pending_hitl. The full suite is 223/223 across 38 files at 71.59% line coverage, with 0 npm audit vulnerabilities. It is deployed EU-resident on Fly.io ams, AMLD6-aligned, with FINTRAC/PCMLTFA mapping.

Abyssal Twin is the edge case: a real-time digital twin on Cloudflare Edge across 40+ regions, with a 52% benchmark cut in token spend via edge telemetry filtering and caching, plus full-fidelity mission replay. The nonprofit client system is the data-governance case: 4 live Looker Studio / LookML semantic panels, SQL query execution time reduced by 30% via CTE refactoring and PostgreSQL indexing, and data sync accuracy driven from 89% to 97%, sustaining 95%+ dashboard accuracy over 8 months for 500+ users.