Open to opportunities

sayan shee

@sayanshee

Message

I build production-ready AI, RAG, and MLOps systems.

India

Message

I'm building a task-driven object-detection pipeline for visually impaired users at IIIT Allahabad, combining Florence-2 with Phi-2/Phi-3, multimodal reasoning, and a Structured Object Knowledge Base to prioritize objects around user-defined tasks.

I architected ResearchGPT, an agentic RAG research-paper assistant using LangGraph, FastAPI, React 19, Qdrant, and LLMs. It routes queries across retrieval, web search, claim verification, and direct answers; its evaluation harness achieved 0.997 answer relevancy and 0.994 faithfulness.

I also built an end-to-end MLOps pipeline for Swiggy delivery-time prediction, achieving 3.16-minute MAE across 45,593 orders. I automated experiment tracking, CI validation, containerization, and AWS EC2 deployment using MLflow, DVC, Docker, GitHub Actions, ECR, and CodeDeploy.

Experience

Work history, roles, and key accomplishments

Current

Swiggy Delivery Time Prediction – MLOps Pipeline

Current

Personal Project

Jan 2026 - Present (9 months)

Orchestrated a 6-stage DVC pipeline for 45,593 food-delivery orders, engineering features and training a stacked ensemble achieving 3.16 min MAE. Implemented MLflow tracking, automated pytest CI gates, and containerized deployment to AWS EC2 via GitHub Actions.

Python DVC MLFlow Docker EC2 GitHub Actions CI CD fastAPI LightGBM Random Forest Pandas NumPy

Current

ResearchGPT – RAG Research Paper Assistant

Current

Personal Project

Jan 2026 - Present (9 months)

Architected an agentic RAG system using LangGraph, FastAPI, React 19, Qdrant, and LLMs, with dynamic query routing and isolated per-session retrieval state. Implemented multi-source ingestion, vector search, SSE streaming, and a DeepEval evaluation harness achieving high relevancy and faithfulness scores.

LangGraph fastAPI React Qdrant LLM RAG Vector Databases Embeddings Reranking DeepEval Python

Current

Research Assistant (Thesis)

Current

Indian Institute of Information Technology, Allahabad

Jan 2025 - Present (1 year 9 months)

Engineered an end-to-end Florence-2 + Phi-2/Phi-3 multimodal AI pipeline for task-driven object detection, integrating MLCoT reasoning with a Structured Object Knowledge Base. Evaluated the system against baseline pipelines for task-aware filtering and assistive computer vision.

Python PyTorch Transformers Multimodal AI Object Detection Computer Vision

Current