I'm a student at IIT Delhi with a deep interest in applied AI, GenAI systems, and product engineering. I don't just study these fields — I build with them. My projects range from deployed full-stack applications with real users to rigorous ML pipelines evaluated against enterprise benchmarks.

I'm drawn to problems where good engineering and thoughtful product design intersect: making AI accessible, building systems that scale securely, and extracting actionable insight from data. I'm actively looking for software engineering and AI/ML internship opportunities where I can contribute meaningfully from day one.

Live at portfoliolite.tech

A premium React Native mobile application for personal net worth tracking — engineered local-first with zero cloud dependency, biometric security, and a hardened payment & license system.

• What it does: Tracks assets across categories with interactive growth timelines, allocation insights, and biometric-protected access. All data stays exclusively on the user's device.
• What makes it technically deep: Built a device-tethered license engine using Supabase Edge Functions (Deno) with HMAC SHA-256 webhook verification to prevent code sharing; engineered a brute-force-resistant parameter parser for inconsistent Razorpay redirect payloads; implemented IP-based rate limiting (5 failed attempts → block) on the edge; automated transactional email delivery via Resend API; DNS-mapped custom domain to GitHub Pages with 100/100 performance score.
• Stack: React Native, Expo SDK 52, TypeScript, Supabase (PostgreSQL + Edge Functions), Razorpay, Resend API, GitHub Pages

Live at hapiimood.me

A production-deployed full-stack web application built to support student mental health through anonymous, AI-driven conversations and mood analytics.

• What it does: Real-time empathetic chat powered by Groq's Llama-3.1, background sentiment analysis, and a personalized analytics dashboard (mood trends, sentiment distribution, weekly engagement).
• What makes it technically deep: Solved a non-trivial Clerk × Supabase RLS identity bridge using Next.js Server Actions; implemented serverless-safe background logging via Vercel's waitUntil; built in-memory IP rate limiting (60 req/min) to protect LLM API costs; configured custom DNS CNAME routing and dynamic CSP headers for production hardening.
• Stack: Next.js 15, TypeScript, Supabase (PostgreSQL + RLS), Clerk Auth, Groq API, Tailwind CSS v4, Framer Motion, Recharts, Vercel

An end-to-end Retrieval-Augmented Generation system benchmarked on the WixQA dataset (6,221 enterprise help articles from Wix AI Research).

• What it does: Full pipeline from knowledge base ingestion → hyperparameter search → retrieval evaluation → generation evaluation → system improvements (reranking, query rewriting, semantic chunking).
• Results: Achieved 100% Context Recall at optimal configuration; reranking improved F1 by +16% on expert queries and +32% on simulated user queries.
• Stack: Python, FAISS, SentenceTransformers (BAAI/bge-base-en-v1.5), Groq (Llama-3.1-8B), LangChain, ROUGE evaluation, LLM-as-judge

A rigorous empirical study comparing 5 prompt engineering strategies across 2 LLMs on the Titanic survival dataset — 10 experimental conditions, 7,000+ real API calls.

• What it does: Evaluates Zero-shot, Few-shot, Chain-of-Thought, Self-Consistency, and Tree-of-Thought prompting against Accuracy, Precision, Recall, and F1 metrics; includes systematic hard-case failure analysis identifying 3 structural LLM reasoning blind spots.
• Key finding: Zero-shot with the 70B model achieved 82.2% accuracy — and complex prompting strategies couldn't overcome strong statistical priors, revealing a fundamental ceiling for prompting-only approaches on tabular tasks.
• Stack: Python, Groq API (Llama-3.1-8B & 3.3-70B), Pandas, scikit-learn, Matplotlib, Seaborn, Jupyter

Live at Study-Buddy

A production-grade AI study companion for students — built around the insight that good tutoring means guiding thinking, not just supplying answers.

• What it does: Ultra-low-latency AI tutoring via Groq inference, personalized study planning, multi-session persistence, and full markdown rendering for code and formulas.
• Engineering highlights: Secure serverless API routing (API keys never exposed client-side); migrated from Gemini to Groq to reduce response latency for a real-time study feel; full light/dark mode with persistent local session history.
• Stack: Next.js 15, React 19, Groq SDK (Llama-3.1-8B), Tailwind CSS, Vercel

A comprehensive data analysis project for a retail business, focused on translating raw sales and inventory data into actionable business decisions.

• What it does: Sales performance analysis, inventory health assessment, logistics optimization, and category-level growth insights — structured as a consulting-style analytical report.
• Stack: Python, Jupyter Notebook, Pandas, data visualization libraries

A responsive weather application with real-time data fetching via third-party weather API integration, location-based lookup, and a clean, intuitive UI.

• What it does: Live weather conditions, temperature, humidity, wind speed, and multi-day forecast for any searched city or detected location — with dynamic UI that responds to weather state.
• Engineering highlights: Demonstrates clean API integration patterns, asynchronous data fetching, error handling for invalid queries, and responsive layout design.
• Stack: HTML, CSS, JavaScript, OpenWeatherMap API

Languages Python · TypeScript · JavaScript · SQL

Frontend & Frameworks Next.js · React · React Native · Expo · Tailwind CSS · Framer Motion

Backend & Infrastructure Supabase · PostgreSQL · Vercel · Clerk Auth · Next.js API Routes

AI / ML Groq API · LLaMA 3.1 · FAISS · SentenceTransformers · LangChain · RAG · Prompt Engineering

Data & Analysis Pandas · Jupyter Notebook · ROUGE · LLM-as-judge evaluation

• Deepening expertise in LLM systems — retrieval, evaluation, and agentic workflows
• Exploring Bioinformatics applications of AI at IIT Delhi
• Expanding Hapiimood's feature set and user base
• Applying advanced prompt engineering strategies (CoT, Self-Consistency, ToT) to tabular and structured data tasks

I'm open to internship opportunities, research collaborations, and conversations about AI systems and product engineering.