The operator behind the review

Jay Sharma.

"The model works" is not the same as "the system is ready." I have spent my career where models, releases, operational ownership, and business outcomes meet - and I run Depilot's reviews personally.

Why I can make this judgment

I built the evaluation and launch-governance system for AI products serving more than 300 million users at Indeed - the gates that decided what shipped and what stopped. Before that, I led a 100+ person product and engineering organization at Amazon Canada, and carried $5B of annual revenue through an ML compliance deadline.

I have shipped AI and ML in marketplaces, fintech, payments, compliance, recommendations, and customer-facing workflows. I have also made the stop call: I switched off a $24M-a-year channel when the system-level evidence showed it was harming marketplace quality. That was a cost saving and a decision-discipline call, not an AI result - and it is the same call a kill verdict asks of your leadership.

I now build and operate an agentic product myself, with explicit human control over consequential decisions. The review method comes from operating these systems, not from studying them.

Operator experience includes - employer experience, not Depilot client results
300M+people served by the AI matching and launch-governance systems built at Indeed
90%reduction in bad matches after LLM + RLHF matching shipped
$5Bof annual revenue carried through the Amazon Canada ML compliance deadline
$24Msaved per year by switching off a legacy channel the evidence said was harming the marketplace

How I work

  • One pilot, one decision, one written verdict. No open-ended engagements.
  • Evidence over narrative: metrics, logs, evals, and real cases traced end to end.
  • The first failed gate stops the review from moving on - that is what keeps it fast.
  • Independent judgment: no implementation contract, no vendor referral fees, no incentive to recommend more build.

Writing

The Failure Notes are where I publish the patterns: one failure mechanism per note, with the warning signs and the evidence that would change the decision. I am also on LinkedIn, or directly at jay [at] depilot [dot] ai.

Bring one pilot and one decision.

In 30 minutes, we identify the decision your team is avoiding and test the five production gates. If a full review is useful, you leave with a clear scope. If it is not, I will say so.

Book a private pilot triage