Data Scientist, Commerce Forecasting

You turn years of measured sale history into the one number a shopper can trust: buy now, or wait.

Product.ai is the verified truth layer for shopping: the intelligence that tells you what is actually true about a product, including when to buy it and when to wait. SimplyCodes is our first proof at scale. Our robots run real checkouts and test discount codes before a shopper trusts them — codes that actually work. It runs at about $22M in revenue and roughly 60% margins. Profitable. Bootstrapped. Founder-owned since 2009. No outside investors. No board. Fewer than twenty operators, outbuilding companies 10x our size.

Why This Role Exists

We already tell shoppers which codes work. The next question is harder: when should I buy? Nobody here owns that answer yet.

Every store has a rhythm. You will read years of measured sale history and tell millions of shoppers the best day to buy. With receipts. That is the Sale Calendar, and it is your seat.

This is a founding individual-contributor role. You direct agents and work beside a handful of elite operators; you are not building a data-science org. Your leverage is judgment: which signal is real, when a forecast is honest, how sure we get to sound. Not headcount. The Sale Calendar ships first inside our certainty work — the same verify-before-you-trust standard behind SimplyCodes. It starts as a pilot across a small set of stores and grows from there.

The System You'll Need to Model

• Sale timing is event forecasting on calendar-driven series. Every merchant runs cycles: holiday resets, end-of-season clearance, monthly promo cadences, one-day flash events. Some rhythms repeat like clockwork; some are noise dressed as a pattern. Separating true cadence from coincidence across thousands of merchants is the core modeling problem: signal versus noise at population scale.
• Calibration is the product, not a diagnostic. A wrong "wait" costs a shopper money and costs us their trust forever, so the loss is asymmetric and reliability is the bar: a "70% chance it drops" that is right about 70% of the time. Raw accuracy matters less than a forecast that means what it says. You are shipping a promise, not a point estimate.
• Most of the catalog is a sparse-series, cold-start problem. A few large stores have dense sale history; the long tail has a handful of observed events per store. The craft is pooling: borrow strength across related series so the model says something useful about a store it has barely seen, and abstain — "not enough history yet" — when honesty demands it.
• The labels are observed outcomes, not proxies. We measured the sales ourselves. This is the moat: verified records of what actually happened at checkout, across years and thousands of merchants — not scraped advertised prices. Anyone who has fought proxy labels knows what that is worth. The forecast is only as good as the history, and ours is ground truth.
• The last mile is uncertainty communication. A model output of "0.72" has to become a sentence a shopper reads and trusts: buy now, wait a week, this is as low as it gets. Turning a predictive distribution into calibrated plain English, without overclaiming, is as much of the job as the model.
• You work inside Cortex, and you model our direction yourself. Cortex is our shared AI brain: the substrate that runs the company and the product family we sell. Every operator works through governed AI sessions, and the substrate answers its own questions from more than 8,600 documents. You will forecast inside it, not beside it — it is how work happens here, and it is a category of one. The company evolves weekly; reading where it is going is part of the seat.

If reading that energizes you, keep going. If it feels overwhelming or underspecified, this isn't the right fit.

What You Will Own

• A consumer-scale probabilistic forecasting system, end to end. Ours is called the Sale Calendar: from raw verified sale history to the shipped buy-or-wait verdict a shopper acts on. You own the models, the backtesting, the calibration, and the judgment on what is ready to show a human — not one layer of the pipeline, the whole line. The craft you must already own: probabilistic time-series modeling, pooling across sparse series, backtests that hold out of sample, Python and SQL end to end. What you grow into here: AI agents as your build capacity, and LLM systems that carry your distributions into consumer language.
• The calibration standard: the honesty bar for every number we print. You define what "confident" is allowed to mean, how we score whether our forecasts held, and when the system abstains because the history is too thin. Reliability measurement, scoring rules, abstention policy — the governance layer every forecasting shop needs, and here it is yours to set. When the calibration bar and the ship date conflict, you hold the bar.
• The decision layer that turns a distribution into a verdict. With our product and design operators, you decide how a forecast becomes a sentence: buy now, wait, this is the floor. The shopper hears exactly as much certainty as the data has earned, and no more. Few forecasters get to own uncertainty communication all the way to the consumer surface; here it is half the job.
• Your seat charter and one number. This is the model we run here: within your first quarter you co-sign a seat charter, a written split of what you decide freely and what you propose for sign-off, anchored to one machine-checkable number that proves the seat works. For this seat that number is calibration: are your buy-or-wait calls right as often as you claim they are. You own it, and it is checkable without anyone's opinion.

Who You Are

You reason from first principles about uncertainty. You know the difference between a model that fits the past and a forecast that holds out of sample, and you can feel when a pattern is real versus when you are overfitting a story. You treat a confidence number as a commitment, and you are comfortable saying "we don't know yet" out loud.

Agents are your production system, and you verify what they hand back. You can build the whole pipeline by hand — the data pulls, the backtest harness, the reliability curves — and that mastery is exactly what lets you trust, or reject, what an agent produces, rather than accept it on faith. You move from a raw question to a shipped, calibrated answer without getting stuck in analysis. The expensive thing here is a redo cycle, never the compute.

You have shipped forecasting or prediction systems that real decisions ran on — live models in front of users or money, not notebooks that never left your laptop. Maybe that was demand forecasting, dynamic pricing, promotion modeling, energy or capacity load, risk scoring, or experimentation; the domain matters less than the discipline. You can show the work: the model, the calibration, and the honest write-up of where it was wrong. We care about the artifact and the reasoning far more than where you did it.

Who this isn't for. This role is wrong if you want to tune models in isolation and hand the "productizing" to someone else — here the model and the sentence a shopper reads are the same job. It's wrong if what you are chasing is a leaderboard rank, a publication, or a brand-name logo to stand on; we hire on the work in front of us, not the pedigree behind it. It's wrong if you want a title and a team to manage more than you want to own the outcome yourself. And it's wrong if you need a fully specified problem handed to you — the buy-moment is under-defined on purpose, and reading the system ahead of the brief is the job. You'll be happiest here if you want to own the whole loop, from history to shipped certainty, and be measured on whether shoppers were right to trust you.

How We Evaluate

We don't run whiteboard interviews or take-home puzzles.

1. Async video screen. Short and self-recorded — about 15 minutes. We want to see how you think, not how you present.
2. Calls with company stakeholders. Short conversations with the operators you would build alongside.
3. Conversation with the founder. How you model uncertainty, where you push back, and whether you can hold the calibration bar in a live argument.
4. Paid work trial. A paid 10-14 day trial — real work in our real environment, taking a live slice of the Sale Calendar from raw sale history to a calibrated buy-or-wait call. It is paid because it respects your time, and because real, paid stakes are the only honest signal. We watch four things: how you get grounded, whether you state your uncertainty before you are asked, how you verify what your agents produce, and whether your self-assessment is honest.

If the work above reads like yours but your resume is unconventional, apply anyway. We hire on the work and the reasoning, not the pedigree. Our evaluation is demonstrated performance on work-relevant tasks — nothing else.

Compensation & Ownership

Total first-year comp: $300,000 – $425,000 (base + performance-based ownership and profit-share programs). Base: $200,000 – $260,000 — top of market for a data scientist at this level.

Beyond base: eligibility for the company's ownership and profit-share programs — grants are performance-based, with terms discussed at the offer stage — plus 100% company-paid family health premiums. Your compute is effectively uncapped, steered by return, never rationed to save tokens.

This is a partnership, not a pay grade. The model is built to mint partners — when the company wins, you win, in real and liquid dollars, every year.

Based in Santa Monica, Los Angeles — in person, five days a week. The rooms are real rooms.