Data & AI systems that make decisions —
and prove they're right.

Designed and built FHI Heat's SQL warehouse from zero — a dbt-style raw → staging → marts model (hundreds of tables/views) unifying ~15 source systems across the US & EU via nightly Python ETL: every commerce, ad, search, ERP, fulfillment, and marketing platform the company runs on. The single source of truth every downstream tool queries.

To replace a third-party creator-intelligence vendor, I built a standalone pipeline that scrapes TikTok's public creator, video, and comment data — the two things no official API returns: creators outside your own shop (for cold affiliate outreach) and raw comment text scored for buy-intent. It feeds the affiliate outreach team and the decision engine below.

Synthetic rebuild, live at the demo below. The desk, the campaign-verdict logic, the closed-loop self-scoring, the RAG copilot & the UX are mine; the creator-ranking model was a colleague’s. Every figure is fabricated.

The operator desk that told FHI's affiliate managers what to do each morning: which TikTok GMV Max campaigns to scale, cut, or tune, when to spark an organic winner, and which creators to sample next. A deterministic rules layer makes every campaign call (ROAS vs. margin breakeven); an LLM copilot only narrates the verdict and cites the audience evidence. It closes its own loop — every campaign recommendation is archived and later graded (difference-in-differences vs. unchanged campaigns) against what actually happened, then tomorrow's actions re-rank by dollar impact. Shipped and in use (a live page + a weekly one-click report). Honest scope: the upstream creator-ranking/scoring model was a colleague's; my work is the decision desk, the campaign-verdict logic, the closed-loop self-scoring, the RAG copilot, and the full operator UX. Because the production system is private, I rebuilt the whole desk public, live, and open-source (demo + repo below) — so you can click through every view and read the code yourself, on fully synthetic data.

The models are only useful if someone acts on them, so I built the alerting layer as an n8n workflow: a daily schedule fans out to four watchers — new high-intent creators in the sampling queue, campaign verdicts flipping, model drift, and the 8–14-day 'spark' window on videos — each polling the deployed engine, diffing against the last run, and alerting only when something changed.

During acquisition diligence, the buyer flagged a large apparent shipping discrepancy — the ERP showed far more paid to the postal carrier than shipping labels were actually bought for. Working solo in the warehouse, I reconciled 570K+ shipping records against the SAP payment ledger and proved it was a bookkeeping mis-classification, not missing money: shipping is prepaid through two carrier 'wallets' that were both being booked under a single vendor. Counting both wallets and both payment streams, spend matched payments to ~0.5% over three years — and the correction was booked.

TikTok pitched its Fulfilled-by-TikTok program (they warehouse & ship your inventory at a higher fee) with claims that faster shipping converts better and gets more traffic. Instead of trusting the brochure, I priced every fee and every claim against FHI's actual first-party data and recommended against blanket adoption — with a per-SKU strategy: don't convert the light-weight brushes (~95% of volume, structurally costlier under FBT), pilot the heavy tools where FBT actually saves, and make TikTok underwrite the unverifiable 'free traffic' claim. Delivered as an interactive dashboard for a non-technical decision-maker.

Below is a public Power BI rebuild on synthetic data — the production dashboard described above was a private Next.js / MSSQL app.

The executive BI web app FHI's CEO and operators ran on — daily revenue, margin, channel, inventory, and ad views where every panel ends in a ranked 'Actions' band (cut / fix / scale / watch) plus an alert layer (stockouts, ROAS collapse, creator decay). It reframed the dashboard from display to decision, and surfaced what the DTC-only reports missed: a major wholesale channel, and a large slice of ad spend with no measurable attribution.

One warehouse the whole company can ask in plain English — from an intern to the CEO — each role scoped to exactly what it's cleared to see, every answer grounded in a single source of truth. A LangGraph agent turns the question into governed SQL (read-only, self-repairing) with role-based access enforced in three layers: a schema-scoped prompt, a guard, and per-role DuckDB views (hidden columns physically absent). Public rebuild on synthetic data modeled on my warehouse; live demo with five role logins (ask it to 'delete all orders' and it refuses).

A retail warehouse modeled in dbt on BigQuery (raw → staging → marts) with data tests, lineage docs, and GitHub Actions CI running the full build + 13 tests on every push. Public rebuild on synthetic data — the named-stack, reproducible version of my production modeling.

SteMy is a hands-free voice AI lab partner for stem-cell researchers. Turning stem cells into heart cells is notoriously irreproducible batch-to-batch — and researchers are hands-busy at the hood during time-sensitive windows, with dozens of variables to track. I build and lead the voice-interaction layer that lets them capture and recall the right experimental variable at the right moment, hands-free. A separate, ongoing UW–Madison research engagement with the People and Robots Lab (Prof. Bilge Mutlu) × the Kamp Lab (Dr. Timothy Kamp); part of an NIH R01 funded by the National Institutes of Health.