Sanchit Pandey

01 / About

Building at the frontier of ML

I'm a final-year student at BITS Pilani focused on LLM research and production AI systems. My first-author paper on RAG utilization failure in sub-7B models is under review at ACL ARR 2026.

On the engineering side, I've built Dual-PPO RLHF systems by modifying TRL internals, production hybrid retrieval pipelines with FAISS + BM25 + LightGBM re-ranking, and full-stack backends in Go and FastAPI that ship in production.

Currently building a multimodal AI platform for India's Ministry of MSME integrating multilingual ASR, OCR, and hybrid retrieval to connect small businesses with government procurement networks.

Tech Stack

ML / AI

PyTorch HuggingFace TRL BitsAndBytes FAISS LangChain scikit-learn AzureML

Languages

Python Go C++ JavaScript TypeScript Dart

Backend & Databases

FastAPI Node.js PostgreSQL MongoDB Redis Firebase

Mobile & Frontend

Flutter React Next.js TailwindCSS

Tools

Docker Git Vercel Linux

02 / Research

First-author publication

Under review · ACL ARR 2026

Can Small Language Models Use What They Retrieve? An Empirical Study of Retrieval Utilization Across Model Scale

Sanchit Pandey · arXiv:2603.11513 · March 2026

Demonstrates a net-negative exact match trade-off for sub-7B models under RAG — showing that simply adding retrieval can hurt smaller models. Introduces a parametric knowledge split that isolates utilization failure from retrieval quality, tested across 5 models (360M–8B), 3 architecture families, and oracle/BM25/dense conditions on 1,000 QA pairs.

5 models

360M–8B range, spanning 3 distinct architecture families

2,588

Oracle failures classified into 6 error categories

61–100%

Irrelevant generation — the dominant failure mode across scales

arXiv Paper Code & Data

03 / Experience

Where I've shipped code

Genpact AI/ML Engineering Intern

Jul 2025 – Dec 2025 · Bengaluru

Adapted the Absolute Zero Reasoner (AZR) from code tasks to NLP by replacing deterministic checks with an LLM-as-a-judge, enabling self-play training for sarcasm detection and NER without labeled data
Achieved ~30% relative improvement within 8 self-play iterations; redesigned AZR to a Dual-PPO setup with custom reward injection by modifying TRL PPO internals to mitigate reward hacking and model collapse
Benchmarked 5 extractive (Lead-3, TF-IDF, SumBasic, TextRank, LexRank) and 1 abstractive (BART) summarization baseline on CNN/DailyMail; evaluated end-to-end pipeline on ROUGE-1/2/L/Lsum
Developed AI systems integrating DeepSeek OCR, hierarchical reasoning, and knowledge graph construction for procurement analytics

Python PyTorch HuggingFace TRL LangChain AzureML Docker

Busy InfoTech / Indiamart Software Engineer Intern

May 2025 – Jul 2025

Delivered production backend endpoints for Pooraa (business management software) using Go, Sails.js, and PostgreSQL
Reduced API latency by 15% across a 5-member team through targeted debugging and streamlined testing workflows

Go Sails.js PostgreSQL Redis Docker

Indian Red Cross Society Data Analyst

May 2024 – Aug 2024

Designed and automated Python ETL pipelines processing 7 years of financial data; performed anomaly detection and data validation, improving allocation efficiency by 25%

Python pandas Jupyter

04 / Projects

Things I've built

★ Featured Project

IndiaAI MSME Intelligent Mapping System

Multimodal matching engine for the Ministry of MSME's TEAM initiative to onboard suppliers to the ONDC network. Decoupled FastAPI backend integrating multilingual ASR, automated OCR form-filling, and a hybrid retrieval pipeline with learning-to-rank re-ranking.

Integrated Whisper ASR for multilingual speech input, DeepSeek OCR for automated form-filling, and a hybrid retrieval pipeline (FAISS + BM25)
LightGBM LambdaRank re-ranker using Haversine geographic distance and semantic similarity scores for buyer-seller mapping
Designed independently deployable components with structured API contracts

Python FastAPI FAISS BM25 LightGBM Whisper Next.js TypeScript

GitHub

Hospital Visitor Management

Full-stack system with QR-based entry, Aadhaar-linked registration, OTP authentication, and automated Excel/PDF reporting.

↓ 70% check-in time · ↓ 90% errors

Node.js React MongoDB Firebase

Live Demo GitHub

AI Selfie Review App

Production mobile app with real-time camera capture, Snapchat-style live filters via DeepAR SDK, and a remote ML prediction backend.

Google Auth · Firestore · ML Pipeline

Flutter Firebase DeepAR SDK REST APIs

E-Bike Rental Platform

Cross-platform Flutter app + Node.js API for campus e-bike rentals with QR booking, OpenStreetMap search, and real-time tracking.

Piloted 5 bikes · 90% utilization

Flutter Node.js MongoDB

GitHub

DeepFake Detection Pipeline

CNN + ResNet50 fine-tuning with Error Level Analysis preprocessing on a mixed synthetic/real dataset. Augmentation, Early Stopping, Dropout, and Batch Normalization.

80.12% test accuracy

Python PyTorch CNN ResNet50

AI Architecture Diagram Generator

Converts natural-language project descriptions into structured XML architecture diagrams using MermaidJS and Draw.io.

↓ 70% manual diagramming effort

Python MermaidJS Draw.io

RhythmCo Music Player

Interactive Flutter music player with collaborative playlists, real-time streaming, shuffle mode, and social likes.

Flutter · Firebase

Flutter Firebase

05 / Beyond Code

Leadership & Achievements

Ethos Professional Development Club

President · Aug 2024 – Jul 2025

Led a 40+ member club organizing professional development workshops and career guidance sessions at BITS Pilani. Mentored 150+ students on internship preparation and technical skill building.

Competitive Programming

Olympiad Qualifier · 2020

Qualified for the Indian National Olympiad for Informatics (INOI) and Zonal Computing Olympiad (ZCO). Holds Google certifications in Network Security and Security Management (SIEM & Playbooks).