AI Systems & Engineering

I build production AI systems — the architecture, the models, and the engineering that makes them dependable enough to ship.

My work sits where modern AI meets real technical and scientific software — where “it runs in a notebook” is the start of the problem, not the end of it. The interesting part is everything after: reliability, evaluation, architecture, and the judgment to know when a system is actually good enough to put in front of people.

Areas of work

Production LLM systems: Designing and orchestrating large language models into features that hold up under real inputs, latency budgets, and cost constraints.
Document & multimodal intelligence: Extracting structured meaning from documents, tables, and mixed media — the unglamorous core of most useful enterprise AI.
Retrieval & knowledge systems: RAG and knowledge architectures that ground models in an organization's own information rather than the open web.
Agentic workflows: Multi-step, tool-using agents scoped to do real work reliably, with explicit boundaries and understood failure modes.
AI evaluation & reliability: Measuring whether a system is genuinely good — and keeping it good as models, data, and requirements shift underneath it.
Scientific AI: Bringing AI into scientific and technical software, where correctness and interpretability are not optional.
Enterprise AI architecture: System design that lets AI features share infrastructure, evaluation, and guardrails instead of each being a bespoke one-off.
Forward-deployed engineering: Working directly with the people who use these systems, and shipping against their real problems rather than a proxy for them.

I keep confidential Datacor and customer details out of public writing. The below describes capabilities and impact, not customer data.

Selected production work

A sample of production AI I’ve led at Datacor. The metrics are from real deployments.

AI cash-application platform: A multimodal document-intelligence system — OCR, layout-aware parsing, and LLM ensembles with judge-model validation — reaching 92%+ schema-level accuracy on complex financial documents, cutting manual review by ~70% and raising end-to-end throughput 8–12×.
AI knowledge copilot: A hybrid RAG + knowledge-graph platform over engineering software and technical documentation, improving retrieval relevance 30–50% and reducing support burden ~40%.
Enterprise AI architecture: A unified data model plus reusable RAG, agent, evaluation, and governance components that let AI features share infrastructure across business units — with organizational standards for LLM evaluation, observability, and human-in-the-loop validation.
Agentic code modernization: AI-assisted refactoring of multi-million-line legacy C++/C# engineering codebases, using a strangler-fig strategy, static-analysis mapping, and CI/CD parity gates to preserve behavior.

Selected public work

A few open repositories that reflect the technical range — scientific computing, clinical NLP, and document intelligence.

DFT-Clustering IGOR Pro Processes DFT calculations of NEXAFS spectra with tensor-based modeling — the tooling behind the Physical Review Letters work.
ChemNER Python A custom named-entity-recognition model for extracting and labeling chemical compounds from text.
PyUMLS_Similarity Python Computes semantic-similarity metrics between concepts in the UMLS medical ontology — clinical NLP infrastructure.
Clinical-Trial-Semantic-Structure-Visualizer Python Extracts and visualizes medical concepts from clinical-trial eligibility criteria.