Headline

AI/Quant @ Zepto | IIT KGP | Author, TowardsDataScience

About

AI/Quant Lead @ Zepto — I build the AI-augmented supply chain behind quick commerce. Moved forecasting, PO, and TO off Excel onto an automated Databricks pipeline, cutting expiry CPO by 70% and lifting availability from 78% → 93%. Lately I've leaned into agentic systems: a Text-to-SQL agent over 20M+ rows/day that decomposes questions and runs its own retrieval, and an LLM + OCR pipeline that autonomously validates and ingests 2,000 invoices/day. Plus XGBoost forecasting and scheduling engines on Genetic Algorithms + MILP under the hood. Previously at Skypoint, I built an agentic Text-to-SQL stack — structured query decomposition into multi-chain retrieval with ColBERT embeddings — going fully OpenAI-free (vLLM + SkyPilot), cutting costs 34% and pushing accuracy from 65% → 90%. Earlier, at Publicis Sapient, I shipped a Virtual Try-On SaaS platform for 100k+ users. IIT Kharagpur grad, comfortable across Python, SQL, TensorFlow/PyTorch, and cloud. Increasingly drawn to pointing frontier AI at the whole sprawl of real-world problems still waiting to be solved.

Location

Bengaluru

AI stack

codex

cursor

Work experience

AI/Quant at Zepto, Dec 2024 - Present · 1 yr 9 mos, Bangalore Urban, Karnataka, India, Development of AI-Augmented Supply Chain: • Built the AI-augmented supply chain stack across FMCG, Fresh and Cafe by migrating PO, TO and Demand Forecasting from Excel-driven workflows to a Databricks-based automated pipeline, reducing expiry CPO by 70% and improving network availability from 78% to 93%. • Led cross-functional teams of data scientists, engineers and PMs to design forecasting systems using XGBoost and PO/TO scheduling engines using Genetic Algorithms and MILP, improving replenishment accuracy and operational efficiency. • Developed a high-throughput Text-to-SQL chatbot over aggregated supply-chain data (20M+ rows/day), enabling real-time analytical insights and reducing RCA turnaround time from days to minutes. • Built an LLM + OCR based invoice extraction system to process 2,000 invoices/day, automating validation and ingestion workflows and reducing PO processing time by 50%.

Co-Founder at Madyam, Dec 2023 - Present · 2 yrs 9 mos, Bengaluru, Karnataka, India

AI/ML Engineer at Foundry, Oct 2024 - Nov 2024 · 2 mos, Bengaluru, Karnataka, India, At Foundry, we‘re not waiting for a decentralised financial future — we‘re building it now. By empowering institutions with the tools they need to mine and stake digital assets, we‘re thoughtfully driving the industry forward. Learn more at foundrydigital.com

Sr. AI Engineer at Skypoint, Dec 2023 - Sep 2024 · 10 mos, Bengaluru, Karnataka, India, Structured Retrieval-Augmented Generation (RAG) via Multi-Chain Text2SQL: • Enhanced the Text-to-SQL vertical of Skypoint Copilot by creating a LLM-agnostic solution using vLLM inference engine and SkyPilot cloud framework, eliminating OpenAI dependence and reducing costs by 34%. • Leveraged Multi-Query RAG through Structured Decomposition of user queries using Mistral-7B-Instruct, followed by retrieval over the database using ColBERT embeddings and dynamic few-shot prompting to reduce the prompt size to under 2,000 tokens. • Improved query accuracy from 65% to 90% through advanced instruction tuning and leveraging state-of-the-art open-source LLMs like the SQLCoder family.

Senior Associate Data Science L1 at Publicis Sapient, Aug 2021 - Nov 2023 · 2 yrs 4 mos, Knowledgebase | Custom Vector Database • Led client engagements and created a custom vector database, a python library with a pluggable interface for Elasticsearch and Weaviate. • Expanded the use-cases including document insertion to create AsktheDocs by making it compatible with OpenAI embeddings and Langchain library • Served the Vector Database over FastAPI and deployed it on a Kubernetes cluster while developing CI/CD using Github Actions and ArgoCD • Ensured the robustness and stability of the system by rigourous testing of python library, FastAPI and API client using pytest to a coverage of 91% Virtual Try-on | Vision-based Saas Platform • Created a Virtual Try-On based SaaS-platform which has varied offerings including Clothes, Hairstyle, Haircolor, Makeup and Jewellery Try-On • Used Bilateral Segmentation Network(BiSeNet) for semantic segmentation of the facial features of humans into 17 different classes • Implemented Parser-Free Appearence Flow Network for on custom clothing dataset generated using Pose estimation and Human Parser Segmentation • Trained custom CNN models using transfer learning for adding pre-validations checks and post-processing enchancement • Used Mediapipe and CVzone libraries for the generation of face mesh, facial landmarks and background removal for real-time image processing • Managed a team of 4 frontend and DevOps engineers to design and develop a feedback system for the SaaS-platform for consumer analytics • Developed CI/CD for the SaaS-platfrom in AWS using Serverless Function, API Gateway for 100k+ users with a max limit of 1000 concurrent users

Education

Bachelor of Technology - BTech, Metallurgical Engineering, Indian Institute of Technology, Kharagpur

Central Board of Secondary Education, Indian Business