Enterprise AI · Agentic Systems · Applied AI Research
I lead the AMI chapter at Capital One and work on AI platforms. As an independent AI researcher, I build original systems from scratch: a trained-and-benchmarked System-1 decision model, and enterprise AI theses adopted as conference talks.
7 sessions across 5 conferences (2026–2027), including 3 solo talks
"The Context Layer: Why Enterprise AI Needs Its Own Operating System"
"Code Quality, Security, and Culture in the AI Coding Era"
Backed by Linux Foundation & Meta
"Closing the Confidence Gap: Making AI Output Trustworthy in Enterprise Systems"
Selected by program committee · ~2,000 enterprise IT executives · Board incl. FedEx, HSBC, Mastercard, Fidelity, Barclays
"INFRASTRUCTURE 1: The AI Data Layer", a panel with peers from Cigna, eBay, and World Wide Technology
"How Are You Using AI?"
"AI-Native Engineering Teams & Production-Ready AI"
"Building Trust in Agentic AI: From Pilot to Production"
Reviewer and program-committee service for ACM, ICSE-co-located, and international workshop venues
Invited by the technical programme chairs to serve as reviewer for the Research Track (long papers); reviews due December 2026, with the rebuttal period and program assembly through January 2027.
Reviewing submissions to the ACM FAccT 2027 technical program, with focus on AI evaluation, assurance, governance, and fairness of automated decision systems.
Appointed to the Junior Program Committee; reviewing technical-debt research submissions alongside the senior PC, with paper bidding in October and reviews in November–December 2026.
Program-committee member; reviewing submissions on generative and agentic AI for software engineering, verification of model outputs, and security of AI-generated code in the December 2026 review cycle.
Independent, from-scratch AI systems research, trained and benchmarked on my own hardware
An original decision model that evaluates AI agent traces - tool calls, tool outputs, final answers - and returns typed judgments with calibrated probabilities. Original decision-head architecture and training pipeline over a BERT-family encoder, trained end-to-end on a MacBook Pro on locally generated synthetic data. Open source under Apache-2.0 with reproducible training and evaluation scripts.
99.7% macro accuracy across 5 agent-trace judgment families on 2,500 held-out test cases - at 24ms latency, $0 per judgment, and bit-exact determinism (0.0 probability deviation over 100 repeated inferences): properties text-generating LLM judges cannot match. Also published the five-layer trust benchmark - schema conformance, calibration under out-of-distribution rules, determinism, perturbation invariance, pinned reproducibility - run against both the model and GPT-class LLM-judge baselines, with honest results either way.
Enterprise AI needs its own operating system: a governed context and retrieval layer between models and enterprise data. Delivered as a solo talk to 12,000+ attendees at AI4 2026.
AI output trustworthiness is an engineering problem solved by layered validation: type conformance, confidence calibration, contract validation, and observability. Solo talk, PyTorch Conference NA 2026 Responsible AI track.
Evaluation systems must themselves be evaluated - for calibration, determinism, and cost. A recursion the industry is converging on through cheap, typed System-1 judge models; subject of forthcoming applied research.
My academic profile, including publications and citation history.
View on Google Scholar →I lead the AMI (golden machine image) chapter at Capital One and work on AI platforms. My focus is the unglamorous middle: making AI-generated work trustworthy enough to run in production.
On the research side, I design, train, and benchmark original AI systems independently - including a from-scratch System-1 typed-decision judge - and author theses on enterprise AI architecture that major conferences have adopted: the Context Layer, the Confidence Gap, and the who-evaluates-the-evaluator recursion.
My foundation is FinTech at scale: a U.S. patent in natural-language rule generation, payments platforms, and AI-powered customer service. Speaking slots at AI4, PyTorch Conference NA, ONUG, and Enterprise AI World in 2026 alone, with an accepted solo talk at Building Business Capability 2027. I serve as a peer reviewer for the ACM FAccT 2027 Conference, on the Junior Program Committee of TechDebt 2027 (ICSE-co-located), and on the program committee of GenSE 2027.
My work sits at the boundary of rigorous computer science research and practical, production-grade software.