I turn complex financial data into products people trust.

Every data question queued behind engineering — and early LLM prototypes made it worse: hallucinated tables, broken joins, answers nobody could trust enough to ship.

Context before generation. The assistant retrieves the right models from DBT metadata — schema, relationships, business definitions — before writing a line of SQL. Golden datasets and an evaluation framework gate every release.

Business users self-serve answers in plain English, and accuracy became a managed metric instead of a hope. The pattern now underpins the platform's AI roadmap.

Bank clients consumed data through fragmented models with duplicated logic and definitions that disagreed — slow onboarding, engineering on the critical path of every report.

A layered modelling strategy — raw, consolidated, business-ready — with the consolidated layer as the primary consumption surface. Models designed around business use cases, not source systems, governed by data contracts and documentation standards.

Reporting sped up, definitions converged, onboarding stopped waiting on engineering — and the marts became the foundation the AI querying layer was built on.

UPI had no standard way to authorise recurring payments — every PSP improvised one-off flows, and there was no reliable way to resolve users across apps.

Treat the ecosystem as the user. Designed the UPI Mandate lifecycle — create, modify, pause, revoke — and the Central Mapper spec, then aligned 10+ banks, fintechs, and TSPs on reference BRDs, PRDs, and TSDs that became the integration standard.

Recurring payments became a rail-level capability used across the UPI ecosystem, shipped alongside compliance work that cut systemic incidents like duplicate debits.

Data platform health was tracked through engineering signals (job runs, tickets, delivery dates), so nobody could answer the product question: are clients actually getting fresh, trustworthy, usable data?

Define success from the consumer's side, not the pipeline's. I set a product-level metric dictionary covering freshness, data-quality pass rate, extract delivery, pipeline health, documentation coverage and test coverage, applied the same way to every client tenant. I deliberately kept delivery tracking out, because program management owns that and mixing the two blurs accountability.

One scorecard shows, for each client tenant and each product, what needs attention now. Platform reviews moved from anecdotes to numbers, and gaps like undocumented models became a visible backlog instead of hidden debt.