Principal Data Analyst, AI-Enabled Analytics at Walmart

I build AI agents that let people talk to their data.

I build the agents, and the governed context underneath that makes their answers trustworthy. The LLM proposes; deterministic code decides.

Monthly revenue in 2025Top 10 customers by total order valueWhich product categories have the most returns?

Opens the live demo. Answers take 20–30 seconds.

  • AI agents
  • Semantic layers
  • MCP servers
  • Text-to-SQL
  • RAG
  • Ontologies
  • Graded evals
  • BigQuery

The flagship, live

An agent that watches a product team's KPIs on live data, and explains them when they move.

(01) Flagship · live

MarketCase

Governed KPI agent · live Polymarket data

A product team's KPIs on live Polymarket trades, defined once in a governed semantic layer, watched every hour and explained when they move. When one leaves its usual range, an agent checks for a data problem first, then breaks the change down by category, market, trade size, trader type and odds band. The model never writes SQL: every sentence cites a receipt carrying its query, and a verifier drops any number it can't trace.

  • 18KPI series checked hourly
  • 28days of same-hour history
  • 9questions answered each morning
Open MarketCase ↗
MarketCase showing Polymarket KPIs checked against their usual range, with an investigation whose numbers link to their queries

How I build

The LLM proposes. Deterministic code decides.

The model writes the SQL; tested, rule-based code decides whether it runs. People trust an answer they can check, so every step leaves a trace.

  • Show the work
  • Retrieval over prompt-stuffing
  • Read-only by design
  • Charts without another model call
Read the case study
Trace · sql-rag live demo

Question

“Top 10 customers by total order value”

  1. (01)

    Retrieve — similar known-good queries from a vector index

  2. (02)

    Schema — only the tables and columns they need

  3. (03)

    Generate — Gemini 2.5 Pro writes BigQuery SQL

    the LLM proposes
    SELECT user_id, SUM(sale_price) AS total
FROM order_items
GROUP BY user_id
ORDER BY total DESC
LIMIT 10
  4. (04)

    Validate — rule-based checks before anything runs

    code decides
    • read-only
    • known tables
    • byte cap
  5. (05)

    Execute — run on BigQuery, chart the result

question → chart in 20–30 sRun this question ↗

What I'm building now

Agents, and the governed context that makes them trustworthy. At Walmart, so the work is internal and there are no links.

  • columns in an ontology and lineage layer

    1,700+ tables, built by parsing 900+ production SQL jobs

  • data requests closed by an autonomous Jira agent

    plus a self-service analytics agent business teams use daily

  • MCP servers with ~75 agent tools

    two drive web apps through a headless browser

  • tests behind graded evals with safety traps

    because an agent nobody trusts doesn't get used

Selected work

Live demos and open source: agents, evals, document AI and the data plumbing underneath.

  1. (01)live demo · rag · text-to-sql

    SQL RAG: text-to-SQL that shows its work

    Retrieval-augmented text-to-SQL on BigQuery. It retrieves similar example queries, writes SQL with Gemini 2.5, checks it, runs it and charts the answer. Every answer shows the examples it used, the tables, the model, the tokens and the BigQuery cost. The production version at Walmart reached 95% SQL accuracy.

    View project ↗

  2. (02)live demo · evals · document AI

    DrawCheck: construction pay-application checker

    Checks a contractor's monthly bill (PDF, Excel, Word or a scan) before anyone pays it: the line math, retainage, change orders and last month's figures, citing the cell behind every problem. Scans are read by a vision model and OCR, and a number counts only when both agree. Caught a real line billed at 105% in a public city bill.

    View project ↗

  3. (03)live · multi-agent · GCP

    LabDrift: keeps Google Cloud labs current

    Reads Google Cloud's daily release notes, finds the course-lab steps they break, drafts a fix, has a second model critique it, and queues it for review. 134 verified fixes on Google's public training repo; deployed with Terraform on Cloud Run.

    View project ↗

  4. (04)streaming · MCP · verification

    Odds arbitrage scanner: sportsbooks vs. prediction markets

    Watches five sportsbooks and Kalshi and Polymarket in real time, checks every sportsbook price against the page it came from, and prices every cross-venue pair after fees. 12 hours of live data, 942k price updates and 1250 pairs say these markets are efficient: the best edge was 80.75¢ per dollar.

    View project ↗

  5. (05)open source · agents · evals

    Interview Prep: a job-search agent that never hits send

    Turns a job posting into a tailored resume, an apply packet and recruiter emails, built as Claude Code skills. It writes drafts and never sends them: outreach waits until I mark the role applied, and every number on a resume must come from a verified claim. I ran my own search on it (126 roles, 50+ applications), then split it into an installable template with behavior evals and a trace that explains each step.

    View project ↗

  6. (06)data engineering

    joinmap: join discovery

    Discovers valid single- and multi-hop joins across unfamiliar BigQuery databases from schema and data evidence, with SQL pushdown.

    View project ↗

  7. (07)rag · best in show

    Proposal RAG for FTI Consulting

    RAG over past proposals, adopted by FTI Consulting: ~30% faster drafting, +42% retrieval with adaptive chunking, +34% generation with query rewriting.

    View project ↗

  8. (08)mcp

    PFF MCP server

    Gives Claude access to Pro Football Focus premium stats through browser automation and first-party JSON APIs.

    View project ↗

  9. (09)full stack

    Prediction-market data workspace

    Kalshi and Odds API ingestion with credit-aware budgeting, FastAPI governance services and role-based React dashboards.

    View project ↗

Want to see more? Everything on GitHub ↗

Where I've done this work.

MS Applied Data Science, University of Chicago (4.0) · BS Information Systems & Economics, Loyola University Chicago

Download resume (PDF)

  1. Walmart

    2024 – now

    Principal Data Analyst, AI-Enabled Analytics

    Internal AI consultant for my org. Built a text-to-SQL copilot over Walmart BigQuery schemas (95% SQL accuracy, ~80% faster query creation), agents that monitor KPIs, and the ontology and MCP tooling underneath.

  2. UChicago Data Science Institute

    2024

    Graduate Student Researcher

    Persona-driven multi-LLM chatbot (OpenAI + Gemini) for a staged donation experiment with 500+ participants.

  3. FTI Consulting

    2024

    Capstone: RAG proposal generator

    Deployed and adopted for client work. Best in Show, UChicago MS in Applied Data Science capstone.

  4. Innovare

    2022 – 2023

    Data Scientist

    Risk model and dashboards across 10+ school districts: +7% retention, −15% course failures.

Let's talk about agents, semantic layers or MCP.

Building something where people need to trust what an AI tells them about their data? I'd like to hear about it.

[email protected]LinkedInGitHub