AI/ML Engineer · Bengaluru, India

Upasana Banerjee

I build multi-agent AI systems that turn complex, messy data into clear decisions — from clinical diagnosis grounded in 68,000+ disease-symptom relations to real-time, multilingual medical reports.

Explore my work Download CV
  • Multi-agent systems
  • GraphRAG
  • LLM agents
  • Multilingual ASR
diagnose.py — clinical pipeline

$ python diagnose.py --patient new

> ingesting EHR ✓

> graphrag 68,000 relations

> differential_dx 12 candidates

> confidence 0.94 ↑

✓ report generated in 4 languages

68K+ disease-symptom relations
4 languages in production
5 AI projects shipped
8.5 B.Tech GPA / 10
Scroll

01 / About

AI engineer with a bias for shipping.

I work where research meets real workflows — designing retrieval, reasoning and speech systems that people can actually use.

At Healtether Health, I helped build a differential-diagnosis system powered by a multi-agent framework and GraphRAG over 68,000 disease-symptom relationships extracted from SNOMED CT, AIIMS Bhubaneswar and 20+ hospitals across Tamil Nadu. I also helped cut the system's false positives by 1.2× through synthetic oversampling of rare diseases.

Later, I built a real-time doctor–patient report generator using Google's medASR model and a medical GraphRAG layer, with multilingual support for Hindi, Bengali, Tamil and Telugu. Before that, at Praxis Business School, I worked across Hadoop, Spark and classical ML pipelines.

Today I'm most at home building agentic systems — tool use, retrieval, evaluation and graph-grounded reasoning — that make complex data explainable and useful.

02 / Experience

Where I've built and learned.

  1. Healtether Health · Chennai, India

    AI Intern

    Jan 2025 — Jan 2026

    • Helped build a clinical diagnostic system for patient differential diagnosis using a multi-agent framework with GraphRAG over 68,000 disease-symptom relationships extracted from SNOMED CT, AIIMS Bhubaneswar and 20+ hospitals in Tamil Nadu.
    • Reduced false positives of the diagnostic system by 1.2× through synthetic oversampling of undersampled rare diseases.
    • Built a doctor–patient real-time report generation system using Google's medASR model and disease-symptom med-GraphRAG, with multilingual support for Hindi, Bengali, Tamil and Telugu.
    • GraphRAG
    • Multi-agent
    • medASR
    • Neo4j
    • LLMs

  2. Praxis Business School · Kolkata, India

    Data Science Intern

    May 2022 — Aug 2022

    • Gained practical knowledge in Hadoop, Big Data and MapReduce.
    • Created a blog post and a Colab notebook demonstrating MapReduce with Python.
    • Implemented machine learning pipelines using Spark (Transformers, Estimators, Logistic Regression).
    • Built complex pipelines combining Logistic Regression, Random Forest and Gradient Boosting.
    • Hadoop
    • Spark
    • MapReduce
    • Scikit-learn

03 / Selected work

Projects & experiments.

Retrieval, agents and applied machine learning — each built to answer a question I couldn't stop thinking about.

01 Computer Vision · RecSys

Personalised Fashion Recommendation System

A deep-learning fashion recommender that suggests clothing from an image input. Uses CNN and transfer learning with ResNet on ImageNet features, plus a content-based filtering model to surface visually similar items.

  • CNN
  • ResNet
  • Transfer learning
  • Content-based filtering
02 Agents · Automation

Agentic Web Browser

An AI-powered browsing agent that accepts a website URL and performs user-defined tasks — research, navigation and form filling — built with smolagents, FastAPI and the Helium headless browser.

  • smolagents
  • FastAPI
  • Helium
  • Tool use
03 LLM Analytics

TicketBot

An end-to-end pipeline for analysing app reviews, using LLM-powered topic discovery, clustering and trend analysis to turn raw feedback into product signals.

  • LLMs
  • Topic discovery
  • Clustering
04 Web Agents · Healthcare

Medicine Price Checker

A medicine price checker for Indian marketplaces that finds cheaper alternatives based on the primary drug component, powered by the TinyFish Web Agent.

  • TinyFish
  • Web agents
  • Scraping
05 RAG · Multimodal

VideoRAG

Retrieval-augmented generation over YouTube videos with structural and semantic filters, using Qdrant as a multimodal embedding store for grounded answers.

  • RAG
  • Qdrant
  • Multimodal embeddings

04 / Toolkit

The tools I reach for.

01

Languages

  • Python
  • C / C++
  • R
02

Backend & Databases

  • FastAPI
  • PostgreSQL
  • Neo4j
  • ChromaDB
03

Machine Learning & AI

  • TensorFlow
  • PyTorch
  • LangChain
  • Hugging Face
  • vLLM
  • scikit-learn
  • smolagents
  • LangGraph
  • OpenAI Agent SDK
  • TRL
04

Cloud & Data Engineering

  • AWS
  • BigQuery
  • Terraform
  • Databricks
05

DevOps

  • Docker
  • Kubernetes
  • GitHub Actions

05 / Education

Foundations and recognition.

B.Tech · 2019 — 2023

Electronics & Communication Engineering

Neotia Institute of Technology Management and Science
West Bengal, India

8.5/10 Cumulative GPA

Hackathons

  • 2nd

    Amazon ML Hackathon 2021

    Hosted on HackerEarth

  • 2nd

    NITMAS Sankalpa Hackathon 2022

    Built an emotional sentiment analysis system from real-time voice transcription.

06 / Contact

Let's build something intelligent.

I'm open to AI/ML engineering roles, applied research and ambitious agentic products. If that sounds like your team, my inbox is open.

ubanerjee2001@gmail.com