Building AI products that deliver real value
A practitioner's portfolio — documenting what I've built, how I think, and the tools I've shipped. From production AI systems at scale to independent experiments in vibe coding.
On-Device · Edge AI · iOS
A growing focus of mine is building AI systems that run entirely on-device — no cloud dependency, no data leaving the phone. These apps demonstrate that meaningful intelligence doesn't require a server: on-device LLMs, edge inference, and adaptive models that learn from user behaviour, locally.
On-Device Photo Culling App — Built with Claude Code
An iOS app that learns how you think about photos — and quietly clears the clutter for you. Every night it surfaces a batch of 50 photos recommended for deletion, learns from every keep/delete decision, and gets smarter over time. All intelligence runs entirely on-device: no cloud, no API calls, no data leaves your phone. A working example of Edge AI and Human-AI Interaction (HAI) in practice — built end-to-end using Claude Code with zero prior Swift or iOS experience.
Built with Claude Code · No prior Swift experience · Runs entirely on-device
Credit Card Decision Engine & Expense Ledger — Built with Claude Code
A fully offline iOS app that tells you which credit card to use before you pay — and lets you log every transaction by voice. Manages multiple cards with different reward rates, spend caps, and milestones. All intelligence runs on-device: on-device STT, an on-device LLM (MLX Swift · Qwen2.5-1.5B) for fuzzy slot extraction, and a deterministic rules engine for reward math. No backend. No data leaves your phone. Built end-to-end using Claude Code with no prior Swift experience.
Built with Claude Code · No prior Swift experience · Fully offline · Video demo coming soon
Independent Projects · Vibe Coded
I build production tools independently using AI-assisted development — no traditional software engineering background. These are live, publicly accessible products: from AI resume optimisers to autonomous job agents to policy modelling engines. Each was taken from idea to deployment entirely by me, using Claude Code, Lovable, and Google Apps Script.
Monsoon Commute Decision Tool · Gurgaon
A monsoon intelligence app that tells Gurgaon commuters whether to take their car or book a cab — before they leave home. Evaluates both legs of the commute independently using real-time rainfall forecasts (Open-Meteo), 83 mapped waterlogging chokepoints across Gurgaon, historical flood depth data at 75th percentile, and the user's car's wading capacity. Returns a single verdict with a confidence score. Built end-to-end with Claude Code, deployed as an installable PWA.
AI Resume Optimiser
Rewrites your resume to match any job description using the Claude API — without fabricating experience. Reframes existing skills in JD language, scores ATS match before and after rewriting, and outputs a formatted DOCX. Built end-to-end with Claude Code and deployed on Vercel.
Autonomous Daily Job Scout
A fully autonomous job search agent that scrapes LinkedIn daily via Apify, scores each listing against a personalised preference profile using Claude, and delivers a ranked shortlist to my inbox at 5PM every day — with one-click apply links. Zero manual effort once deployed. Built on Claude Desktop MCP + GitHub Actions cron.
Shipper Bidding Marketplace
A real-time transport procurement marketplace connecting shippers and transport providers. Features a gamified live auction mechanic where carriers bid on freight loads — making price discovery transparent and competitive. Built as a take-home for Delhivery's VP of AIML Products.
Policy Modelling Tool
An interactive analytical tool that lets users explore how India's Lok Sabha seats could be redistributed using factors beyond raw population. Born from the real policy debate after the defeat of the 131st Constitutional Amendment Bill (April 2026). Users distribute 100% weightage across six factors — population, GSDP, HDI, geographic complexity, SC/ST share, fiscal dependency — and see live seat projections for 23 states with full calculation breakdowns.
Adaptive Motivation Engine for Sales
A CRM overlay that adapts motivational nudges and task prioritisation based on a sales rep's real-time performance patterns. Designed to reduce motivation drop-off and decision fatigue in high-churn sales environments — serving the right kind of push at the right moment using behavioural signals.
Production · Policybazaar
A three-phase recommendation engine that personalises learning paths based on user background, intent, and observed behaviour. Progressed from rule-based cohorting → deterministic preference capture → Bayesian regression — building toward adaptive, data-driven suggestions as data maturity increased.
A domain-bounded AI chatbot for life insurance discovery — reducing confusion in a high-involvement, low-trust category. Designed a curated vector knowledge base, a guardrailed RAG pipeline (retrieval → validation → generation), and an LLM layer optimised for clarity and regulatory safety — not persuasive selling.
An AI-powered voice and chat assistant embedded in the insurance proposal form journey. Combines speech-to-text, real-time translation, tokenisation, and a RAG backend to guide customers through complex form-filling — reducing drop-off at the highest-friction touchpoint.
A multi-model AI learning system that adapts content and difficulty based on observed user behaviour. Designed to reduce discovery friction and improve knowledge retention for insurance products — serving both customer-facing and internal agent education use cases.
Background
I'm a Product Manager focused on building products that solve real user problems and drive measurable business impact. Currently at Policybazaar Insurance Brokers, where I own the AI product stack — from ML recommendation engines to RAG chatbots to AI voice systems.
I work at the intersection of users, data, and stakeholders, with a deliberate focus on applying AI where it earns its place — reducing friction, improving decisions, and creating outcomes that wouldn't exist without it.
Beyond my day job, I build tools independently using AI-assisted development. This site is a living record of that work — not a pitch deck, but a repository of what I've shipped and how I think.
Product Thinking
Zomato optimises for price, ratings, and speed. A growing urban segment wants healthier choices but has no in-app support — forcing them to use external calorie tools, breaking the ordering flow and causing decision fatigue and reduced confidence.
Urban, repeat users. Moderately health-conscious — not strict dieters. Motivated by awareness, not rigid tracking. Everyday decision-makers, not fitness extremists.
Let's connect
I'm always open to conversations about AI products, interesting problems, and what's being built. Whether you're a recruiter, a builder, or someone who stumbled onto the delimitation engine — feel free to reach out.