I build AI products where mistakes are expensive.

Context

Regulated onboarding requires matching a live face to a government ID. Too loose and you onboard fraudsters; too strict and you lose genuine customers at the last step.

My role

Product Owner for the selfie & liveness flow — capture UX, scoring thresholds, and the audit trail an auditor actually reads.

Key decision

Score and keep every attempt instead of a single pass/fail. Retakes became evidence rather than friction, and each decision got a reviewable trail.

Result

2L+ verifications a day, each with a liveness score, a face-match confidence and a timestamped attempt history.

The tradeoff: every point I moved the liveness threshold traded fraud caught against genuine users rejected. A stricter bar blocks more spoofs and also blocks a real customer in poor light holding a cheap phone. I tuned for the lender's loss appetite, not for the cleanest demo — and made the threshold configurable rather than baked in.

Context

Retailers buying fresh produce daily had no formal credit history, so banks wouldn't lend. Working capital was the constraint on how much they could buy.

My role

Product Manager, 0→1 — from the first underwriting rule to a live lending product at national scale.

Key decision

Underwrite on observed buying behaviour on the platform rather than on financial statements nobody had. Transaction history became the credit signal.

Result

~$350M in annual disbursals, with NPAs down from 6% to 2% as the scoring tightened.

The tradeoff: lower NPAs and higher GMV pull in opposite directions. Tightening limits is the fastest way to clean up a book and the fastest way to starve good retailers of the credit that grows it. I moved limits down for thin-history buyers and up for proven ones, so the book got safer without the platform losing its best customers.

Context

Onboarding is a funnel with a dozen places to quit — a failed selfie, a missing document, a bank's OTP timing out. Every abandoned journey is acquisition spend already paid for.

My role

Product Owner — instrumenting the funnel, defining the drop-off taxonomy, and designing the nudge logic on top of it.

Key decision

Treat the reason for the drop-off as the trigger, not the drop-off itself. Someone blocked by a document problem needs a different message from someone who simply got distracted.

Result

Stalled journeys get a targeted nudge tied to the step that failed, instead of a generic "complete your KYC" blast.

The tradeoff: nudges have a hard ceiling before they become spam, and the channel is shared with the lender's own messaging. Retargeting harder lifts recovery and burns goodwill and SMS budget at the same time — so the cap on messages per journey was a product decision, not a technical one.