Technical Product Manager at Times Internet (Economic Times). I ship production AI products (LLM personalisation, voice AI, newsroom AI) and lead a cross-functional AI team, with hands-on engineering depth.
Portfolio · LinkedIn · Email · New Delhi, India
At Times Internet:
- Personalised home feed. LLM content enrichment, user-affinity scoring and rule-based ranking. Per-article enrichment cost down 80% and LLM card latency down 74%. Live for the first user cohorts.
- Voice-AI calling platform. Speech recognition, LLM dialogue and speech synthesis on GPU inference, with sub-700 ms voice-to-voice latency. Built on 36K+ real call transcripts; it tripled outbound call volume.
- Newsroom AI-copy detection. Two zero-shot detectors running in shadow mode, each at no more than 1% false flags.
- Subscription lead engine. Two calibrated gradient-boosted models, for never-subscribed and lapsed readers, that produce one ranked lead list every day.
- Content operations. A clipper that cuts 30-90 s clips from 7-8 hour masterclass recordings, an image-tagging API built on a CLIP, BLIP and DETR ensemble, and AI automation that cut manual ops effort by 40%.
- Team. I lead a cross-functional AI team (hiring, sprint planning, design and code reviews) and led the architecture of its campaign voice agents, AI video studio and audience-data pipelines.
This work lives in private company repositories.
Outside work, I am the founder and product lead of R2C, an IIIT-Delhi research-to-commercialisation platform built by 20+ student and research contributors. I architected the platform and built its social distribution engine.
| Project | What it is |
|---|---|
| finance-rag-system | Question answering over Indian company filings with deterministic fact extraction, source citations and number-verification modules |
| cohortx-task-2 | CohortX challenge, 3rd of 9 validated teams: clinical-trial eligibility criteria to semantic triples with MMR-retrieved few-shot prompting, with the full ablation trail |
| Cement-Concall-RAG | Retrieval and guidance reports over the earnings-call transcripts of 13 Indian cement companies |
| indian-stock-backtesting | Backtesting for Indian equities: 14-metric fundamental scoring, technical-indicator screens and Flask dashboards |
| classify-image-objects | Zero-shot image classification API: CLIP and OpenCLIP ensemble with BLIP captions and DETR object crops |
More: Article-Summary-Insights (four-agent LLM report on a financial news article) · web-automation-framework (browser automation modules from a project with 500K+ automated form submissions) · user-chatbot-langgraph (LangGraph FAQ chatbot) · rocket-flight-controller (model-rocket flight computer firmware)
- B.Tech in Computer Science and Artificial Intelligence, Minor in Entrepreneurship, IIIT-Delhi (2021-2025)
- ML Engineering Intern, Scrap Uncle (2025): multimodal damage assessment and a voice chatbot
- Research Intern, SBI Labs, IIIT-Delhi (2023-2024): survival analysis on the Framingham dataset
- Research Intern, ECE Labs, IIIT-Delhi (2023): greeting bot with 85% intent accuracy
Tools I use most: Python, PyTorch, Hugging Face Transformers, LangChain and LangGraph, FastAPI, Flask, scikit-learn, XGBoost, pandas, React with TypeScript.