Principal Product Manager — ad serving, retail media, and AI evaluation.
At Walmart I own the ad decisioning stack — targeting, pacing, relevance, measurement — and the human-in-the-loop evaluation platform that keeps its models honest. Before that: Amazon, Wish, iHerb. Cornell MBA, CS undergrad, MA in cognitive science.
I don't write specs to argue a point. I build the thing and let it argue.
Everything below runs in a browser. Nothing to install.
LLM Eval Scorecard — ▶ open it Blind side-by-side human scoring of two model responses against a weighted rubric. Randomized pane order so position bias can't leak in, bootstrap confidence intervals on the margin, inter-rater reliability across annotators, CSV in and out for real multi-rater studies.
The bet: an eval you can't audit is a vote, not a measurement.
Adaptation Radar — ▶ open it A composite score built entirely from live public signals — Wikipedia, Open Library, Hacker News, Google Books — with no backend, no API key, and no pre-baked dataset. Ranks by the slope of attention, not the level.
The bet: for anything you're trying to catch early, the derivative beats the absolute.
Ad Creative Optimizer — ▶ open it Predicts creative fatigue across Google, Meta, TikTok, Pinterest and LinkedIn with an exponential-decay model over per-platform performance feeds, then alerts before the decay shows up in ROAS.
The bet: creative fatigue is a forecastable curve, not a postmortem.
Off-Site Ads — ▶ open it The full off-site retail-media flow — objective, budget, bid type, audience, multi-placement preview, seven-day KPI readout — modeled side by side across Amazon DSP and Walmart Connect.
The bet: bid-type selection, not budget, is where self-serve advertisers actually fall off.
Enterprise Intelligence — ▶ open it React + Express prototype that turns enterprise search from a place you go into something that comes to you: what changed, where value moved, where to act — with provenance on every claim, confidence recalibrated server-side from feedback, and agents that draft but never ship without a human.
The bet: a proactive layer only earns trust if every push carries its receipts.
Money Next Steps — ▶ open it Cross-vertical personal-finance recommendations that sequence rather than rank — credit, lending and investing ordered by what unlocks what, with a PM lens that exposes the reasoning behind every recommendation.
The bet: in compounding domains, order of operations beats relevance score.
Sotto — signed macOS and Windows builds, CI green Push-to-talk dictation that inserts text wherever your cursor is. One hotkey per language, speak-to-translate, polish modes, and a personal dictionary for the proper nouns every ASR model gets wrong.
MeetingScribe — macOS menu bar Notices you're on a Zoom, Meet, Teams or Webex call, transcribes it fully on-device, and writes a summary with follow-ups to disk. Nothing leaves the machine.
AI Trackers — three scheduled agents, zero application code
Each tracker is a natural-language SKILL.md file. They gather live sources and deliver a bilingual digest to my calendar on a schedule.
Four procedural animatics from Romance of the Three Kingdoms — Red Cliffs · Guandu · Three Visits · Six Expeditions — every frame drawn in code, each exporting a real video file. Built to find out how far procedural animation could go before it needed an artist.
LinkedIn · chloetan.cornell@gmail.com · San Jose, CA