AI & Data Portfolio

The thread that runs through everything I build. AI is now table stakes, not a feature; data is the moat. This page pulls the AI/ML and data work out of my case studies so you can see it in one place — what I've shipped in production, how I think about feedback loops, and how I decide which AI use cases are worth building.

Production AI I've Actually Shipped

Not prototypes — systems that ran at scale with real accuracy and business outcomes.

Product	What the AI did	Outcome	Where it lives
AI4Bharat / IIT Madras — multilingual LLM	Indic-language model eval + data pipeline; PM-defined quality signals for 22 languages	Accuracy 55% → 93%	Experience deep-dive
ShareChat / Moj — vernacular feed	Ranking & localisation quality signals across 15 Indic languages, 150M+ users	D1 retention +15pp	Experience deep-dive
Agentic Insurance Claims	Multi-agent document pipeline (OCR → extraction → decision), HITL escalation	~$5/claim unit economics	Case study
SastaSundar — pharmacy	OCR prescription extraction + overnight queue automation	Cart-cancellation recovery	Case study

How I Think About AI in Products

Three questions I ask before greenlighting any AI/ML feature:

Does it remove real friction? — If a rules engine or good UX solves it, I don't reach for a model. AI earns its place only where the problem is genuinely fuzzy, high-variance, or unscalable by humans.
Does it improve with data? — A feature that doesn't get better as usage grows isn't an AI moat; it's a one-time cost. I design for the feedback loop first: what signal do we capture, how does it re-train or re-rank, what's the latency from event to improvement.
Does it stay explainable? — Especially in regulated domains (insurance, health), I keep a human-in-the-loop path and a reason-code trail. Black-box decisions that can't be audited are a liability, not an asset.

Data as the Moat

My default mental model for any product:

Event capture → Feature store / signals → Model or decision engine → Product surface → New events
        ↑________________________ feedback loop ________________________↓

Capture before you need it. The cost of instrumenting late is a cold-start you can't undo. I define the event taxonomy at spec time, not after launch.
Proprietary signal > bigger model. The defensible part is the labelled, domain-specific data you own — not the foundation model everyone can rent.
Guardrails are part of the loop. Quality, safety, and drift monitoring are first-class metrics, not afterthoughts.

How I Prioritise AI Use Cases

I score candidate use cases on a simple 2×2 before they ever reach a roadmap:

	High data leverage	Low data leverage
High friction removed	✅ Build now (e.g. claims triage, vernacular ranking)	⚠️ Buy/integrate, don't build
Low friction removed	🔄 Capture data, revisit later	❌ Don't build

Each item above links to the full case study or experience deep-dive elsewhere in this portfolio — this page is the index, not a duplicate.