AI/ML Engineer · Open to opportunities

Decoding complexity.
Delivering clarity

Wireframe data visualization with Insights, Predictions, Impact labels
Mritunjay Pandey
About Me

Mritunjay Pandey

Role Worked
Data Scientist, AI Engineer, ML Engineer, SFT & RLHF Prompt Engineer
Relevant Experience in AI
1.8+ years
Education
National Institute of Technology (NIT) Jalandhar (B.Tech, 2025)
Skills
Supervised & Unsupervised Machine Learning, Deep Learning, Natural Language Processing, Sequence Models (Transformers), LLMs, RAG, (SQL, Pandas, PyTorch, FAISS, Azure ML, Hugging Face)
Strong understanding of fundamentals with mathematics.
Work Experience

My professional journey.

Aditya Birla Group

Data Scientist - GET

Aditya Birla GroupMumbai, India
Aug 2025 — Present
Highlights
  • Assigned to architect a Fashion Recommendation System for ABFRL's planned D2C platform expansion designed for AI-driven personalization, semantic search, and cold-start discovery across an 826K-item catalog.
  • Fine-tuned a Siamese MPNet bi-encoder with Multiple Negatives Ranking Loss (MNRL) on 50K contrastive pairs derived from 2.5M+ customer interactions, improving Recall@10 by 153.8% over BM25 baseline.
  • Optimized production inference using ONNX INT8 quantization and FAISS HNSW, cutting latency by 81% (32 ms to 6.1 ms), reducing model size 4x, and achieving 163 QPS with efficient ANN retrieval.
  • Deployed the system on Hugging Face Spaces with A/B benchmarking dashboard, sustaining end-to-end latency under 50 ms SLA. Currently serving as a working prototype for ABFRL's D2C recommendation layer.
Featured Projects

A selection of my recent work.

Document Q&A RAG Chatbot

Document Q&A RAG Chatbot

RAG · NLP

Hybrid retrieval (ChromaDB + BM25 + Neo4j) over 1,526 textbook chunks. Streamlit chatbot on Llama 3.1 8B.

Llama 3.1ChromaDBBM25Neo4jStreamlit
Multi-Agent Marketplace Automation

Multi-Agent Marketplace Automation

Agents · LLM

Multi AI agent automating dispute resolution, support triage & health monitoring with HITL GPT-4 workflows.

GPT-4Gmail APISlackFlask
Satellite Crop Stress Detection

Satellite Crop Stress Detection

Remote Sensing · ML

GEE + Sentinel-2 pipeline over 500 hectares. NDVI/NDRE/NDMI features with Random Forest Model.

Random ForestPython
High Performance Two-Tower Recommendation System

High Performance Two-Tower Recommendation System

Retrieval · MLOps

RecSys with fine-tuned Transformers Bi-Encoder, HNSW ANN search, and ONNX Int8 quantization for 5x faster CPU inference.

TransformersFAISSONNX
Agent Portfolio and Valuation Advisor (APVA)

Agent Portfolio and Valuation Advisor (APVA)

Agentic AI · LLM

LangChain agent with multi-LLMs inference, 5+ API integrations for Multi-tier pricing of domain.

Llama
Multi-Agent LinkedIn Optimization System

Multi-Agent LinkedIn Optimization System

Conversational AI

AI Agent for profile scoring, job-fit assessment, ATS content optimization and guidance with session memory.

StreamlitConversational AI
Medicine Recommendation System

Medicine Recommendation System

Healthcare · ML

ML assistant for drug recommendations from symptoms & vitals using Kaggle clinical datasets.

Jupyterscikit-learnPandas
Virtual Painting Studio

Virtual Painting Studio

Computer Vision

Gesture-driven drawing surface follows index-finger paths, toggles colors/eraser with multi-finger poses.

OpenCVMediaPipeFlask
My Skills

The stack I build with.

Technologies & tools I work with

Machine Learning

05
Supervised & Unsupervised LearningDeep LearningNLPTime SeriesTransformers (BERT, GPT)

Generative AI

08
LLMs (GPT)LLaMAMistralRAG PipelinesFine-Tuning (LoRA)LangChainAI AgentsVector Databases (FAISS)

Toolkit

09
PythonSQLPyTorchPandasScikit-LearnHugging FaceAzure MLGitStreamlit
Research Paper

Domain-Adaptive and Scalable Dense Retrieval for Content-Based Recommendation

Pandey, M. · Journal of the ACM (JACM), submitted/under review [Manuscript ID: JACM-2026-002]

DOI: https://doi.org/10.48550/arXiv.2602.00899