Work

Systems and open-source workflows, each ending with the trade-off it made and what I would build differently now.

Production · 2024 — present

Talk to Data

“Fifty analysts were filing tickets to ask questions about their own data.”

LangGraph
ReAct agents
Dremio MCP
mem0
PostgreSQL
MLflow
Azure ML
Kubernetes
React / TypeScript

50+

Analysts

US and Europe, daily use

95%

Accuracy

500-question benchmark

$0.05–0.20

Cost / query

tracked per query in MLflow

85%

Consistency

run-to-run, same question

Research · 2024 — 2025

Autonomous Web Agent

“The model could see the page. It could not tell you which box was the submit button.”

LangChain
OpenAI
MCP
RAG
DOM / XPath resolution

30% → 59–68%

Component ID

GPT-4-mini, before and after grounding

MLDS 2025

Publication

first author, peer-reviewed

Shipped · 2024

Stratified Anomaly Detection

“Compare every market to every other market and the small ones all look like errors.”

Isolation Forest
Z-screening
Rolling price windows
Power BI
Python

10–15%

Records screened

every Michelin sales pipeline

unique values required per stratum

Shipped · October 2026

Evidence-first Agent Supervisor

“An agent needs to remember what happened, especially when its tools stop answering.”

Agent Skills
Markdown playbooks
Evidence ledgers
Capability adapters

Single writer

Account ownership

Architectural rule in the published skill, not a performance measurement

Human review

Final submission

Explicit boundary in the published workflow, not an automated completion claim