Final-year CS engineering student. I build the systems around the model — retrieval pipelines, payment-recovery loops, redaction engines — because that's usually what decides whether an AI feature survives contact with production.

Most of what I ship sits at the RAG / automation boundary: chunking strategy, vector-store choice, HMAC-verified webhooks, background schedulers that keep running when the network doesn't cooperate. Two internships so far have both come back to the same question — what happens on the failure path: a scanned document instead of digital, a declined card, a face that isn't in the known set.

📍 Delhi, India · 🎓 B.Tech CSE, Manipal University Jaipur (2023–2027)

RazorShop · live at razorshop.app — a full e-commerce platform where the AI layer's job is to claw back revenue the checkout flow would otherwise lose. A background scheduler watches for failed payments and abandoned carts, opens a recovery case, and has Groq generate an outreach email tailored to the actual failure reason (declined card vs. timeout vs. insufficient funds) rather than a generic template; customers can accept recovery, promise to pay later, or opt out, and every response feeds the recovered-revenue metrics on the merchant dashboard.

On top of that sits a full multi-tenant marketplace: a Razorpay-integrated storefront with cart, addresses, and order tracking; a merchant portal for catalog, inventory, and fulfillment; and an admin portal that reviews seller applications and provisions merchant accounts on approval. Returns run through a 5-stage logistics pipeline (pickup scheduled → picked up → in transit → returned to seller → refund initiated), and stock is only restored once the item is confirmed back with the seller, not the moment a return is requested — a small detail that avoids overselling inventory that hasn't actually come back yet.

The merchant-facing AI assistant takes text or voice input (Groq Whisper for speech-to-text, Sarvam for text-to-speech) to run stock updates, deal creation, and refunds, but nothing high-impact executes without an explicit confirmation step — the assistant proposes the action, the merchant approves it. Razorpay webhooks are verified with raw-body HMAC-SHA256 before anything downstream trusts them, and the whole thing runs on AWS behind Nginx with GitHub Actions handling CI/CD and automatic rollback if a deploy's health check fails. React TypeScript Express PostgreSQL/TypeORM Razorpay Groq AWS

Document Intelligence RAG — PDF Q&A that doesn't answer past what it can point to. Text extraction falls back to Tesseract OCR when a page is scanned rather than digital; chunks get embedded with Sentence Transformers into ChromaDB, and every answer from the Groq-backed pipeline carries a page-level citation back to source. FastAPI ChromaDB LangChain Groq Next.js

Aegis — real-time redaction for live video. YOLOv8-nano flags people and devices, EasyOCR + regex catch visible PII, and a rule-based engine scores each frame LOW/MEDIUM/HIGH — unknown faces get an emoji overlay, devices get blurred, all at ~20 FPS / ~150ms latency on CPU. YOLOv8 OpenCV EasyOCR FastAPI React

FloatChat AI — natural-language access to real ARGO ocean-float data. Descriptive questions go through ChromaDB semantic search; analytical ones get translated NL→SQL. Ingests the actual global GDAC dataset into Postgres at 100K+ profile scale, and runs against local (Ollama) or cloud LLMs interchangeably. Reached the internal round of Smart India Hackathon 2025, out of 400+ teams. FastAPI Streamlit PostgreSQL ChromaDB Ollama/Groq

FreelanceX — two-sided freelance marketplace, one API serving both a Next.js web client and a native Kotlin Android client. Full order lifecycle from pending to completed, JWT-based client/freelancer roles, and a 5-star review system underneath it. Next.js Express MongoDB Kotlin

IT Intern, IFFCO-Tokio General Insurance — Jun–Aug 2026, Gurugram RAG pipeline over 500+ page insurance policies; 9 FastAPI endpoints; automated underwriting reports, ~40% faster than manual review.

Automation Intern, HCLTech — Jun–Aug 2025, Noida Document automation across Azure Document Intelligence + Power Automate, ~20 fields extracted per PDF, PDF→Excel pipeline end to end.