AI Product thinker & Data Analyst. 3+ years building GenAI-powered products from 0→1. I combine behavioral analytics, LLM capabilities, and user research to ship features that move engagement, retention, and outcomes.
I started in Mechanical Engineering at Delhi Technological University — not because I wanted to become a mechanical engineer, but because complex systems with real-world constraints were the most interesting problems I could find. That instinct didn't change. It just found a better domain.
AI products, to me, are an engineering problem with human variables. You zoom out to understand the user's mental model, identify the highest-leverage friction point, form a hypothesis, then ship fast and measure honestly. I've done that at Chegg — building a 3-layer AI recommendation engine from scratch — and at Wokelo AI, shaping a GenAI product roadmap from beta to enterprise adoption.
I'm drawn to problems where the solution has to hold for millions of people — where every point of drop-off, every moment of confusion, every missed recommendation is a person who didn't come back.
Every metric came from deliberate strategy, rigorous experimentation, and relentless execution — across Chegg, Wokelo AI, and beyond. Built on behavioral analytics, NLP, A/B testing, and AI product thinking.
👇 Tap each role to explore
A 3-layer hybrid AI personalisation engine built for Chegg Skills — solving "catalog overwhelm" by guiding students from unclear intent to the right course in seconds, with a structured path toward their career goal.
From bank churn dashboards to mobility subscription models — each structured as a PM case study.
Banks lose high-value customers silently — by the time churn is detected, it's too late. Identify at-risk users before they leave and surface that intelligence to non-technical stakeholders.
End-to-end churn prediction pipeline: feature engineering from transactions, tenure, and product-usage breadth; Random Forest + Logistic Regression classifiers; live Power BI dashboard with risk segments and retention opportunity sizing.
Product-usage breadth was a stronger churn predictor than tenure — counter to the team's intuition. The dashboard made this actionable without requiring SQL access.
Chegg had 30K+ raw user feedback entries with no structured mechanism to surface pain points or actionable product signals at scale.
Full NLP pipeline — TF-IDF for keyword extraction, LDA for topic modelling, VADER for sentiment scoring per cluster. Outputs synthesised into a prioritised product backlog with evidence-backed recommendations.
Sentiment alone misleads. Topic modelling separated surface sentiment from the true product signal — "content is too hard" is engagement, not complaint.
67% of users browse 5+ minutes without ordering. 34% bounce with no order placed — decision fatigue costing session revenue and retention.
Designed an AI Meal Planning Assistant — mood NLP input, weekly planner, budget filter, health mode. Full PM case study: 3 personas, before/after journey, 6-metric success framework, 3-phase GTM (MVP → Premium ₹99/mo).
Weekly Active Planners (habit frequency) predicted D30 retention better than single-session conversion rate.
~23% cancellation rate, 4–7 min acceptance delays, 1.5x–2.5x surge pricing make Rapido unreliable for daily commuters — eroding trust and driving churn.
3-tier subscription model (₹799/₹1,499/₹2,299) with Surge Lock, Priority Matching Engine, Cancellation Shield, Predictive Scheduling, and Driver Incentive Layer. Full PM case study: personas, competitor gap, 6 metrics with baselines, 3-phase GTM.
Driver incentive alignment is the actual moat — not the pricing tiers. Two-sided marketplace subscriptions succeed or fail on supply-side economics.
Generic career advice fails individual skill gaps and job market signals — leaving users with vague, non-actionable guidance they can't act on.
Domain-specific RAG system: FAISS vector search + Hugging Face LLMs. Integrated O*NET skill frameworks, resume parsing, and live job data into a multi-source retrieval pipeline. Explainability-first design with a dedicated /explain endpoint.
Without retrieval, LLMs confidently hallucinate job requirements. FAISS grounding made outputs auditable — and trustworthy.
Existing studies ignored demographic variation, urban/rural differences, and social behaviour as combined predictors of alcohol consumption in students.
Python-based EDA on 649 student records from Portuguese secondary schools. Statistical visualisation (Seaborn, Matplotlib), correlation heatmaps, hypothesis testing across demographic, academic, and social dimensions. Published ICIMMI 2022.
Non-linear dynamics are missed by simple models. The peak appeared at famrel=2, not famrel=1 — strong relationships reduce consumption, but the pattern isn't monotonic.
What makes this rare: I think in product strategy and first principles — then I build the analytics infrastructure myself.
I'm looking for Product Analyst and PM roles where data, AI, and user empathy converge. If you're working on something ambitious — I'd love to hear about it.
Happy to chat about product, AI, analytics, or swap notes on interesting problems. Hit Send message to send directly, or use any of the links below.