I turn models into products people rely on.
Six GenAI systems in production at Stanza Living — from a voice agent handling 40K leads a month to a computer vision compliance loop running across 22 cities. IIM Indore. BIT Mesra.
Systems I've shipped
Production AI that moved real metrics — not prototypes on a slide.
Riya Voice Platform
AI Voice Agent · Lead Conversion
The Problem
Stanza Living's inbound volume hit 70K leads a month — far beyond what any sales team could manage. Response latency was killing conversion, and agents were burning out on repetitive qualification calls.
What I Did
Defined the product architecture for an LLM-powered voice agent with real-time language-drift detection, since our leads code-switch between Hindi and English mid-sentence. Built confidence-score routing so calls showing uncertainty or frustration hand off to a human — the model doesn't try to recover what it can't handle.
The Outcome
Riya now covers the full top-of-funnel autonomously. The team's bandwidth shifted from call volume to actually closing deals.
Impact Metrics
Principles I won't ship without
Hard-won from building AI in production, not from reading about it.
Evals before code
The first thing I do on any new AI feature isn't write a spec — it's build an evaluation set. I define what 'good' looks like in measurable terms before a single line of application code is written. This keeps teams from shipping vibes instead of systems.
Confidence routing
I design every AI system with an explicit fallback path. The 65% confidence gate in our CV audit system isn't a footnote — it's a core product decision. Models handle what they can; humans handle what they can't. That boundary is where most AI products fail.
Security as a day-one requirement
PII masking, prompt injection defense, and latency contracts live in the architecture doc — not the QA backlog. At PFC, regulated data access wasn't optional; it shaped every design choice from the start.
Drift is a product bug
Production models degrade quietly. I found this out the hard way when our CV system's accuracy dipped after a real-world lighting change across clusters. Now I treat drift detection as a product requirement, not an engineering afterthought — with alert thresholds, retraining triggers, and structured human feedback loops built in.
Career Timeline
AI Product Manager
Built the AI product function from scratch — voice agents, computer vision compliance, NL2SQL, generative comms, and an AI video pipeline. Every system is in production and handles real users, real data, and real operational stakes.
Assistant Product Manager
NLP-driven early-risk detection on a ₹40,000 Cr infrastructure loan portfolio. Worked inside a tightly regulated data environment — PII masking and role-based access weren't optional constraints, they shaped the architecture.
MBA
Top-ranked business school. Focused on product strategy and technology management. PM case competition finalist at Tekion and WITDA. Earned PSPO II and CSM certifications.
Co-founder & Product Lead
Built and launched an adaptive-learning platform for competitive exam prep from scratch. Led an 11-person team through MVP and go-to-market. Grew revenue 50% and engagement 40% before transitioning to IIM.
B.Tech, Electronics & Communication
National Talent Search Examination (NTSE) Scholar — awarded to the top 0.1% of students nationally. Competitive engineering foundation that informs how I think about system design and constraints.
A 90-second walkthrough
The short version of how I think about building AI products — straight from me.
Things I build for myself
Side projects where I get to be the PM, designer, and engineer all at once.
Moodbite
Eat for how you feel
A mood-first food discovery app that maps emotional state to what you should eat — with a recommendation loop that learns your taste over time. Built because most food apps optimise for ratings, not for how you actually feel at 10pm.
Featurfit
Ship the right features
A lightweight prioritisation tool for PMs. Input your backlog, score each item on impact and effort, and Featurfit surfaces the highest-leverage bets in order. Built after watching teams spend two hours in a room arguing about stack-rank.
Send_me
Frictionless file handoffs
No sign-up. No cloud storage account. Generate a link, enter a code on the other device, and your file is there. Built for the specific annoyance of needing to move something quickly between two machines you control.