Graham Lee, Senior Technical Program Manager, AI Platforms at Product.ai
Senior Technical Program Manager, AI Platforms

Graham Lee

Senior Technical Program Manager, AI Platforms · Product.ai

The Commerce Truth Scientist. Bets on the human against the hype machine — building the systems that hand shoppers verified truth instead of a promise broken in the fine print. Checkout codes are where it starts; the method generalizes anywhere people are sold hype instead of facts.

Turns raw checkout signal into a measurable win for the shopper — proven on live traffic, not asserted. An ad-tech systems engineer turned commerce-intelligence architect who documents the spine, ships the code, and drives the number until it’s the right number. The north star is a metric that can’t be gamed without betraying the shopper. Because if the code actually works, both sides win.

LinkedIn → product.ai →
Role
Senior Technical Program Manager, AI Platforms
Based
Los Angeles, CA
Studied
UC San Diego — Physics & Economics, Music Composition
Prior
System1 — LA ad-tech / YouTube
1.25M
Checkout codes scored every day on the live verification pipeline
~35
Production systems mapped for the platform migration
20%+
Top-line revenue lift from an ML system built earlier in his career
2
Company missions owned end-to-end
About

The engineer who decides what 1.25 million checkout codes do every day — and refuses to let a single one lie to a shopper.

1.25M
codes scored per day
2
company missions owned
~35
systems mapped for migration
20%+
revenue lift, prior ML build
The range

01 · Architecture altitude

Systems Architecture

Specs the spine before the code — the schemas, contracts, and data lineage a verification platform stands on. The architecture is documented before the first line is written.

02 · Sharpened at Product.ai

ML Modeling

Trained the models that predict whether a checkout code will work — beating the prior approach on both accuracy and calibration, then locking a production-grade candidate.

03 · Production engineering

Pipeline Engineering

Took the scoring pipeline live and kept it live through a run of failure modes in a single day. The fix was the infrastructure that prevents the whole class of failure — not four patches.

04 · Sharpened at Product.ai

Experiment Design

Designs the live A/B experiments that prove verified intelligence moves shopper outcomes — measured on real users, over real windows. Validation, not vibes.

05 · 15-year discipline

Data & Measurement

The empirical instinct, forged across a career in ad-tech: point-in-time correctness, reproducibility, and a refusal to operate against a number he hasn’t verified.

06 · Off the matrix

Cross-System Migration

Mapped ~35 production systems into one dependency picture to plan a platform cutover — every hazard ranked, every surface accounted for before anything moves.

Marquee stops

An ad-tech systems engineer who became a commerce-intelligence architect.

2025 → Now · Product.ai
Product.ai
Senior Technical Program Manager, AI Platforms · Commerce Truth Scientist
Owns checkout-code intelligence end-to-end and drives the architecture of the next-generation commerce-verification platform. Took the code-scoring pipeline live in production — now scoring ~1.25 million codes a day — and brought the first true success-rate baseline into existence by finding and fixing a bug that had been under-counting results by more than half. Documents the spine, ships the code, drives the number.
Commerce Truth Scientist
Before Product.ai
Earlier · Product.ai (Demand.io era)
Demand.io → Product.ai
Senior Technical Program Manager, AI Platforms
Worked across the core backend, data-engineering, and data-science teams — led critical database migrations, internal tooling, and infrastructure, and product-owned the publishing platform across multiple engineering teams. This is where the bridge from ad-tech systems thinking to commerce-intelligence architecture got built — colleagues named the pattern: the strategic thinking of a product manager with the deep technical expertise of a seasoned architect.
Backend · Data · DS span
Early career · System1 (LA ad-tech / YouTube)
System1
Senior Technical Program Manager
Designed and built the experimentation platform that scaled hypothesis testing for paid-advertising optimization across System1’s owned-and-operated sites and YouTube — researching and validating automation across Google Ads, Facebook Ads, and Taboola with supervised and unsupervised machine learning. Collaborated on an ML recommendations system that increased top-line revenue of core business lines by over 20% when productionized.
+20% revenue lift
Before System1 · earliest chapter
Yield Management & Lead Quality
Statistical analysis · model + algorithm deployment
Built trust with business partners by developing yield-management algorithms that differentiated lead quality — predicting likelihood-to-convert on the long tail. That let partners manage their costs, while optimizing yield and growing accounts. The same shape as everything since: empirical prediction that makes both sides win.
Empirical foundations
UC San Diego
University of California, San Diego
BS in Physics and Economics · Music Composition (minor)
Physics, economics, and music — and the combination is not incidental. Physics and economics gave him the measurement instinct; a jazz sensibility gave him the operating posture — shared language, implicit alignment on what we’re after, and a willingness to jam.
Physics × Econ × Music
The covenant

Three generations of medicine, music and jazz from UCSD, and a covenant to truth.

Three generations of Graham’s family have been in medicine. What he inherited was not the profession — it was a conviction: that data and design exist to produce clarity for people in distress.

A shopper at checkout is a small version of that. They want the code to work. They want the price to be fair. They want the transaction clean. Validated intelligence means they get it.

That is why the success rate is the right metric to be measured against — it cannot be gamed without betraying the shopper. If the code actually works, both sides win. The covenant runs all the way down to the number Graham will let himself be judged by.

A code is deployed to win the click, then broken in the fine print. I build the forensics engine that drags that fine print into the light — so the shopper gets the truth, not the hype. Graham, on the work
Operating code

Six principles that show up across his architecture, his code, and his sessions.

01 Method

Spec-first, then fluid.

Architecture documented before code — the schemas, the contracts, the spine. Once the spine is clear, move fast and iterate.

“I document architecture before I touch code. Once the spine is clear, I move fast and iterate.”

02 Posture

Adversarial synthesis.

Explore the problem and the solution space through several models in parallel, then synthesize — covering ground without sacrificing rigor.

“I run the core questions through multiple models in parallel, then synthesize. It’s how I cover ground without losing rigor.”

03 Infrastructure

The meta-fix beats the patch.

When a system has known failure modes, the load-bearing fix is the infrastructure that prevents recurrence — not the one-off hotfix. Build the thing that catches the whole class of failure.

“The meta-fix breaks the find-it, patch-it, repeat cycle.”

04 Standards

Infrastructure-grade on every project.

Point-in-time correctness, typed schemas, reproducibility. No corners cut on the parts that matter — even working solo.

“I don’t cut corners on the parts that matter, even solo.”

05 Scope

No zombie optimization.

No new features built on the legacy stack — triage only, to keep the lights on. Every hour spent on the old machine is a tax paid to buy time for the new one.

“Every hour spent on the legacy stack is a tax paid to buy time for the new machinery.”

06 Covenant

Refuse work that corrodes the user.

The systems must be grounded in truth and beauty — real outcomes for real people. Not dashboards for their own sake. Not metrics that flatter the org.

“If the code actually works, both sides win.”

What he owns

Two missions in flight. One verification spine underneath.

Most of the work is the infrastructure behind the surface. What it adds up to: a verification engine that closes the loop from raw checkout signal to a measurable win for the shopper.

Checkout-Code Intelligence

1.25M / day · live

The pipeline that scores whether a code will work — in production, at scale.

Owns it end-to-end: validating the data, training the predictive model, designing the live shopper experiment, and the scoring service itself. Now scoring roughly 1.25 million codes every day.

Next-Gen Verification Platform

architect of record

The migration to a sovereign commerce-verification platform.

Drives the architecture and the build — schema design, shadow validation against live traffic, and intelligent routing. The systems the rest of the platform reads its ground truth from.

The Success-Rate Baseline

first true measure

The number the whole team now operates against.

It exists because Graham refused to accept the wrong one — finding and fixing a bug that had been under-counting results by more than half. A metric that can’t be gamed without betraying the shopper.

The Migration Map

~35 surfaces

One dependency picture for the entire platform cutover.

Mapped ~35 production systems, ranked every hazard, and grouped surfaces by migration risk — the plan that lets the cutover happen without breaking what shoppers rely on.

What sets him apart

Combinations rare individually. Unusual to find in one operator.

Architect and implementer at three altitudes

A platform specification one day, a schema the next, a production fix the day after. Most operators sit at one tier; Graham moves across all three in the same week.

The empirical baseline-finder

Diagnosed and fixed a bug that had under-counted results by more than half while measuring the first true baseline. The number the team now operates against exists because he refused the wrong one.

The meta-fix instinct

When a run of failure modes surfaced in a single day, the response was infrastructure that prevents the whole class of failure — not a stack of one-off patches.

Refuses metrics that corrode users

The covenant runs all the way down to which metrics he picks. The success rate is the right north star precisely because it cannot be gamed without betraying the shopper.

Ad-tech systems thinker turned commerce-intelligence architect

Years of supervised and unsupervised ML for ad pacing produced the systems instinct; years across publishing, data, and data science produced the cross-layer fluency. Both forged the architect.

Econ measurement, jazz tempo

Physics and economics gave him the measurement instinct; music — orchestral and jazz — gave him the operating posture: play in the same key, hit the changes together, trust the room.

Career history is publicly verifiable; the production results described here are measured on Product.ai’s live systems. See the record →
The through-line

Document the spine, then ship the code that proves it on real users. Refuse work that corrodes the shopper. Drive measurement until the number you stake your name on is the right number.

Product.ai builds with operators like Graham — engineers who own the spine and the shipped result both. See open roles →

Get in touch with Graham See the team