Systems at scale, across the whole revenue function.
Built out the foundation of Databricks' Partner Strategy & Ops team: forecasting, attribution, incentives, quotas, planning — and AI deployed on top of all of it.
The operating spine of a GTM org. I run the forecast, attribution, incentive, and quota systems a partner business depends on, and align them across a two-sided marketplace.
This page ships with its own eval suite. Repo →
Systems that didn't exist before, built to last and run across a two-sided marketplace: the partner team, the sellers who co-sell through partners, and the partners themselves.
Established the cadence the business plans on and built the numbers behind it. Still the operating rhythm today.
The single standard for how partner revenue is counted and credited — the number behind compensation, investment, and executive reporting.
Turned partner incentives into growth levers: programs tied to new consumption and new-logo acquisition, still in force.
Wrote how the field and sales teams engage with partners — unblocking the channel and driving more partnerships, with lead flow in days instead of weeks.
Put partner reporting on autopilot: agents that keep the org current without anyone pulling a report.
Points sellers to the right partner for the deal in front of them, in plain language.
Made the data worth building on: one foundation, shaped with stakeholders, that every agent and insight runs from.
The discipline, demonstrated from scratch: AI authors the rules, tested code owns every dollar.
Stood up at Databricks partner ops and still running: reporting, partner recommendations, and natural-language answers, from raw GTM data to agents reps actually use.
The operating rule, everywhere: the money path never touches a model. Plain-English rules → deterministic, tested code → evals pinning the math.
This repo demonstrates the same discipline, built from scratch on synthetic data: five layers, from governed data to a loop that runs the QBR on the tooling itself.
This isn't a toy domain: incentive crediting is a class of problem I know from running it — and everything here is built from scratch on synthetic data. The rule that makes AI safe in production: the model authors rules in plain English; deterministic, tested code computes anything that touches money. An eval suite pins the math — mid-quarter hires, territory handoffs, split credit, coverage gaps — so no model ever computes a credited dollar, and no deal is silently zeroed.
One résumé, the whole record. The ATS-safe version is single-column for application portals; the default is designed for reading.
The stories behind these systems are better told live.