I'm Pooja Sharma. Customers rarely ask for what they actually need — so I go find it in failure logs, support tickets, and honest conversations, let data settle the debates, and ship. I build AI-powered identity, fraud, and onboarding products that 75+ enterprises run on.
The research, the decisions, the trade-offs, and what actually happened.
The concept: generative AI broke the assumption that the face on camera is real. Rather than one model, we shipped four complementary defense layers — a deepfake classifier, metadata analysis, image-pattern analysis, and injection-attack detection — embedded invisibly inside existing liveness flows.
The concept: a legacy, fixed-workflow product rebuilt as a configurable enterprise platform — no-code controls, configurable webhooks and failure handling, full backward compatibility. Customers never felt the migration happen; they just felt it work better.
The concept: a production-critical KYC capability migrated from AWS Rekognition to a custom model in 3 months. The bet that paid off: data quality over model architecture — production failure audits fed the training set, and a phased pilot → validation → global rollout de-risked the switch.
The concept: evaluated 3+ eSign vendors and implemented intelligent vendor switching — every transaction routed for cost and reliability, with automatic failover so provider outages never become stuck customer journeys.
The concept: a large share of VKYC failures were fixable image problems discovered too late. This API moves validation upfront to capture time — users fix a blurry or cropped document in seconds instead of failing in review and starting over.
The concept: loan applications die in forms. This voice-first agent turns them into a conversation — extracting application data from natural speech and driving KYC and loan-origination-system actions through governed, audited tool calls. The LLM converses; a policy layer owns everything regulated.
Independent deep dives on products I use — because judgment needs practice outside your own domain.
The concept: navigation apps demand active attention, but commuting is passive. Smart Arrival Alerts let commuters choose when and how to be notified — by time, stops, or distance — so the phone goes back in the pocket and the app speaks up only when it matters.
The concept: quick-commerce platforms are built entirely around the sender — recipients can't decline unwanted deliveries used for harassment or unwanted contact. The shift: consent-driven deliveries — accept/decline before fulfilment, block senders, trusted circles for frictionless gifting, and a safety mode for repeat attempts.
The concept: safety isn't a single feature — trust must be built before, during, and after every ride. Safe Ride Mode activates automatically: live trip sharing, smart route monitoring with gentle check-ins, one-tap SOS — plus trust scores and hygiene signals before a rider ever books.
The concept: prevention over redistribution. Surplus Deals let restaurants sell end-of-day inventory at a discount; Cancelled Order Rescue turns fresh cancelled meals into time-boxed flash deals — every stakeholder better off, waste intercepted before it happens.
The concept: answer a business question with no data and defensible logic. A segmented bottom-up decomposition — daily, weekly, and monthly user cohorts with distinct transaction behavior — then a reality check against public NPCI benchmarks.
A role, an idea, feedback on my work, or a good product conversation over coffee.