Innocean India Worldwide Communication Pvt Ltd
AI Intern · Gurgaon, India
Jul 2026 – Present
Building AI tools for brands like Hyundai and Kajaria: ad creation, competitor tracking and a “see the tile in your room” app.
More detail
- Architected an agentic ad-campaign generator for Hyundai: LLM-driven concept generation feeding an image-synthesis pipeline, with identity-lock prompt engineering that cut hallucination artifacts by 30% while preserving brand fidelity.
- Developing a visualizer agent for Kajaria that lets users upload their own floor/room images and preview 60+ tile SKUs with realistic 2D-to-3D visualization.
- Built an autonomous MarTech discovery agent (FastAPI + Next.js) that uses weighted rubric scoring across India's startup ecosystem for real-time discovery, validation and scoring of partners and competitors.
Projects: ScoutINNtelligence · AdINNtelligence · Creative Intelligence Engine · Kajaria Tile Visualizer
IIT Jodhpur
AI/ML Research Intern · Remote
Jun 2026 – Dec 2026
Building a study companion for UPSC aspirants that works out which topics matter most, in Hindi, English, Bengali and Tamil.
More detail
- Building an AI study platform for UPSC aspirants that analyses previous-year questions (PYQs), General Awareness material and the latest news to surface the questions and topics most likely to matter.
- Transcribe educator video explanations (Hindi/Hinglish) with Whisper and align them to news articles using multilingual sentence embeddings, marking important topics alongside the supporting news snippets and transcript excerpts.
- Distilling an educator's teaching style into a reusable style guide that conditions LLM-generated explanations, with resources in Hindi, English, Bengali and Tamil.
- Designing an interactive LLM interface where students can explore a topic and discuss related topics further.
CDIS, IIT Kanpur
AI/ML Developer Intern · Kanpur, India
May 2025 – Feb 2026
Built AI that reads and sorts citizen complaints for government portals, cutting manual review by 60%.
More detail
- Orchestrated end-to-end NLP/LLM pipelines for government grievance platforms, reducing manual review time by 60% through automated complaint summarization, classification and semantic routing.
- Engineered hybrid semantic-lexical search with contextual LLM re-ranking for precedent discovery, reducing case resolution time by 45% while maintaining 95%+ relevance.
Projects: ES-RAG Grievance Retrieval · Evidence Analysis
IIT Kanpur
Intern · Kanpur, India
Dec 2024 – Jan 2025
Built a system that recognises which drone is in the sky from a single radar reading, reaching 96.6% accuracy.
More detail
- Built a radar-based drone identification system from measured radar cross-section (RCS) data for 17 drone and object signatures (DJI M100, Mavic, Phantom 4 Pro, Parrot, F450, Hexa, Y600, Walkera, a mini helicopter and a battery) across HH/HV/VH/VV polarizations, in the 26–40 GHz band.
- Fitted cubic splines through each signature along both viewing angles (phi and theta), so a single radar reading at any angle can be compared with every known drone.
- Identified the drone as the closest-matching spline, cross-checking the phi-based and theta-based matches with a vote, and flagged readings far outside the known range as unknown. Reached 96.6% accuracy in drone-type prediction.
Projects: Radar Drone Classification
I build production LLM agents, retrieval-augmented generation workflows and agentic pipelines, and I've shipped them into government, legal-tech, education and enterprise MarTech settings.
At Innocean India I build the INN agents and generative tools for brands like Hyundai and Kajaria. In parallel, as a remote research intern at IIT Jodhpur, I work on an AI study platform for UPSC aspirants built on PYQs, current news and educator video transcripts. Before that, at CDIS, IIT Kanpur, I built grievance pipelines for government portals: complaint summarization, classification and routing, and hybrid Elasticsearch + vector retrieval with LLM re-ranking.
Much of my work lives in private repositories. Those projects are listed on the home page and marked private; the public ones link to their code.
For the technically curious.