Hi, I'm Subhadeep 👋
Senior Data Scientist and AI/ML Engineer. I love building things and helping people.
SM

About

AI/ML engineer with a background in Computer Science Engineering and over 3+ YoE in the AI/ML industry. My interests span Applied Machine Learning and Deep Learning, Agentic AI, Generative AI, RLHF and MLOps / AgentOps. I have published several papers at IEEE and ICBAI conferences and filed 4 patents in applied generative and agentic AI and reasoning, including a multi-tool agentic speech and audio intelligence framework and a GRPO and memory-based reinforcement learning framework for upgrading small language models in tool-using agentic AI. I enjoy working at the intersection of generative multimodal models, deep reinforcement learning, and applied agentic systems, especially where research has a clear path to production and measurable impact on latency, cost, accuracy, and safety.

Work Experience

A

Axtria Ingenious Insights

Lead AI EngineerBengaluru, India
    G

    Genpact LLC.

    Senior Data Scientist/AI EngineerBengaluru, India
    • Deep Learning and Generative AI R&D in Healthcare and Finance: Generated data synthetically using proprietary and open-source LLMs for fine-tuning LLMs and VLMs for finance, insurance and medical use-cases. Fine-tuned Phi 3 and Phi 3.5 using the synthetically generated datasets. Used Unsloth for faster and memory efficient fine-tuning. Evaluated the models on test-set and created comprehensive report for reference.
    • Multimodal Retrieval-Augmented Generation for QA Systems: Build multi-modal RAG application for question answering on documents (scanned PDFs).
    • Enterprise GenAI and MLOps Solutions in Healthcare and Finance: Deployed fine-tuned models on cloud services. Built consumable REST APIs serving endpoints to run inference on the deployed models.
    • State-of-the-Art Audio Language Frameworks Implementation: Built end-to-end plug-and-play framework and application for audio based use-cases involving summarization, sentiment analysis, insights and feedback generation.
    • Developing Low Code Frameworks for Large Vision Models Fine-Tuning: Developed a library for low code fine-tuning traditional computer vision models using custom data, with minimal code. Supports image classification, segmentation, object detection, image generation and VLM fine-tuning.
    • Papers and patents authoring: Authored in two papers over audio use-cases for IIMB ICBAI-2024 conference. Currently focused on patenting our latest RnD solutions on Generative and multi-modal AI.
    G

    Genpact LLC.

    Data Science InernBengaluru, India
    • Developed state-of-the-art vision transformers for medical and industrial applications. Created explainable AI solutions to visualise and interpret model performance.
    • Designed and built a Flask web application with an HTML user interface to efficiently utilise the model and provide real-time results. Conducted experiments with MONAI Deploy SDK for model deployment in the medical domain.
    • Generated a comprehensive report comparing models based on key metrics such as performance, accuracy, training time, and inference time.
    • Explored vision-language models tailored for medical applications.
    • Developed smart agent based solution with Langchain to automate Doc QnA, Source Retrieval and Summarization with RAG, as a part of Genpact's Inhouse Hackathon.
    • Explored few Multimodal VLMs for Visual Question Answering, focusing medical domain.
    • Developed model for detecting AI generated text. Developed Gradio and FastAPI based User Interface utilizing the models.
    • Experimented with Nvidia's RIVA framework and OOTB models for S2T, T2S and NMT tasks. Created a comprehensive report comparing performance of Nvidia's and open source audio models.
    i

    iNeuron.ai

    Data Science InternRemote
    • Built a stock price prediction model using Machine Learning. Developed the pipeline using Machine Learning algorithms like Decision Tree and XGBoost
    • Built a sentiment classification system using NLP and Logistic Regression, using twitter data.
    I

    Indian Institute of Technology, Bombay

    Flutter Developer InternRemote

      Skills

      Python
      React
      PyTorch
      FastAPI
      HuggingFace
      Flask
      Langchain
      Azure
      AWS
      GCP
      Github
      TypeScript
      JavaScript
      PostgreSQL
      Docker
      Kubernetes
      Java
      My Projects

      Check out my latest work

      I've worked on a variety of projects, from simple websites to complex web applications. Here are a few of my favorites.

      Clinical trials matching with DeepAgents

      Clinical trials matching with DeepAgents

      Agentic AI for Clinical Trials matching, using Decision Model, Deep Agents, HITL and Graph knowledgebase for faster automated trial ranking, matching and auditable neuro-symbolic, verifiable scoring.

      Python
      Deepagents
      Neo4j
      FastAPI
      TypeScript
      PostgreSQL
      TailwindCSS
      Arxplorer - Agent powered Arxiv papers explorer

      Arxplorer - Agent powered Arxiv papers explorer

      Your AI-powered research companion for exploring, saving, and understanding arXiv papers.

      Python
      Langgraph
      FastAPI
      TypeScript
      PostgreSQL
      TailwindCSS
      Qwen Mental Health Chatbot System

      Qwen Mental Health Chatbot System

      Finetuned the Qwen 3 4B Thinking model using an open-sourced Mental Health dataset from Huggingface. Built a FastAPI application to wrap the fine-tuned model with performant inference server and memory to persist user conversations. Added Patient, Cases and User handling to create an End-to-end mental health chatbot application.

      Python
      PyTorch
      HuggingFace
      FastAPI
      TypeScript
      PostgreSQL
      TailwindCSS
      Godot RAG Assistant

      Godot RAG Assistant

      Developed an agentic RAG application for Godot Documentation QnA Assiatent, to help Godot developers to quickly learn/debug games. Uses godot docs, Huggingface Godot qna dataset and other internet sources in real-time to answer queries. Built a simple UI using Chainlit. Currently working on building a proper UI using NextJS.

      Python
      PyTorch
      Langchain
      HuggingFace
      FastAPI
      TypeScript
      PostgreSQL
      TailwindCSS
      Next.js
      Gemini Mail Assistant

      Gemini Mail Assistant

      A Gemini app that summarizes your daily mails from Gmail.

      Python
      Langchain
      HuggingFace
      FastAPI
      Certifications

      Verified Skills

      • Aug 2026
        A

        Claude Certified Associate - Foundations

        Anthropic
      • Sep 2024
        D

        Fully Automated MLOps

        Datacamp
      • Sep 2024
        D

        MLOps Concepts

        Datacamp
      • Sep 2024
        D

        MLOps Deployment and Life Cycling

        Datacamp
      • Jun 2022
        D

        Data Scientist with Python

        Datacamp
      • Jul 2022
        D

        Data Manupulation with Python

        Datacamp
      • Jul 2023
        J

        Deep Learning with PyTorch: Zero to GANs

        Jovian
      • Mar 2023
        J

        Machine Learning with Python: Zero to GBMs

        Jovian
      • Apr 2021
        G

        Google Cloud Training - Cloud Engineering Fundamentals, Cloud Application Development, Cloud ML-AI Google Cloud Training - Cloud Engineering Fundamentals, Cloud Application Development

        Google Cloud
      • Jun 2020
        C

        Java Programming: Problem Solving with Software

        Coursera
      Articles

      Things I've written

      Tutorials and write-ups on RAG, LLM apps and MLOps.

      Building a Multi-Vector Chatbot with LangChain, Milvus, and Cohere

      In the fast-growing area of digital healthcare, medical chatbots are becoming an important tool for improving patient care and providing quick, reliable information. This article explains how to build a medical chatbot that uses multiple vectorstores. It focuses on creating a chatbot that can understand medical reports uploaded by users and give answers based on the information in these reports. Additionally, this chatbot uses another vectorstore filled with conversations between doctors and patients about different medical issues. This approach allows the chatbot to have a wide range of medical knowledge and patient interaction examples, helping it give personalized and relevant answers to user questions. The goal of this article is to offer developers and healthcare professionals a clear guide on how to develop a medical chatbot that can be a helpful resource for patients looking for information and advice based on their own health reports and concerns.

      Self Hosting RAG Applications On Edge Devices with Langchain and Ollama–Part II

      In the second part of our series on building a RAG application on a Raspberry Pi, we’ll expand on the foundation we laid in the first part, where we created and tested the core pipeline. In the first part, we created the core pipeline and tested it to ensure everything worked as expected. Now, we’re going to take things a step further by building a FastAPI application to serve our RAG pipeline and creating a Reflex app to give users a simple and interactive way to access it. This part will guide you through setting up the FastAPI back-end, designing the front-end with Reflex, and getting everything up and running on your Raspberry Pi. By the end, you’ll have a complete, working application that’s ready for real-world use.

      Self Hosting RAG Applications On Edge Devices with Langchain and Ollama – Part I

      This article follows that journey, showing how to transform this small device into a capable tool for smart document processing. We’ll guide you through setting up the Raspberry Pi, installing the needed software, and building a system to handle document ingestion and QnA tasks. By the end, you’ll see how even the smallest tech gadgets can achieve impressive results with a bit of creativity and effort.

      RAG Application using Cohere Command-R and Rerank – Part 2

      In the previous article, we experimented with Cohere’s Command-R model and Rerank model to generate responses and rerank doc sources. We have implemented a simple RAG pipeline using them to generate responses to user’s questions on ingested documents. However, what we have implemented is very simple and unsuitable for the general user, as it has no user interface to interact with the chatbot directly. In this article, we will modularize the codebase for easy interpretation and scaling and build a Streamlit application that will serve as an interface to interact with the RAG pipeline. The interface will be a chatbot interface that the user can use to interact with it. So, we will implement an additional memory component within the application, allowing users to ask follow-up queries on previous responses.

      RAG Application with Cohere Command-R and Rerank – Part 1

      The Retrieval-Augmented Generation approach combines LLMs with a retrieval system to improve response quality. However, inaccurate retrieval can lead to sub-optimal responses. Cohere’s re-ranker model enhances this process by evaluating and ordering search results based on contextual relevance, improving accuracy and saving time for specific information seekers. This article provides a guide on implementing Cohere command re-ranker model for document re-ranking, comparing its effectiveness with and without the re-ranker. It uses a pipeline to demonstrate both scenarios, providing insights into how the re-ranker model can streamline information retrieval and improve search tasks.

      A Beginner’s Guide to Evaluating RAG Pipelines Using RAGAS

      In the ever-evolving landscape of machine learning and artificial intelligence, the development of language model applications, particularly Retrieval Augmented Generation (RAG) systems, is becoming increasingly sophisticated. However, the real challenge surfaces not during the initial creation but in the ongoing maintenance and enhancement of these applications. This is where RAGAS—an evaluation library dedicated to providing metrics for RAG pipelines—comes into play. This article will explore the RAGAS library and teach you how to use it to evaluate RAG pipelines.

      Papers & Patents

      Research and IP

      Peer-reviewed papers and patent filings from R&D work.

      paper

      An E2E Framework to Improve Agentic AI Using GRPO Reasoning and Continuous Memory Based Learning

      ICECET 2026Subhadeep Mandal, Sukanya Bag, Shalini Mittal, Ayushman Gupta, Chirag Jain

      This work proposes a reinforcement learning-based framework for fine-tuning and optimizing Small Language Models (SLMs) used in Agentic AI systems, with a focus on improving tool invocation and decision-making over complex, multi-step tasks. The framework targets both open-source and proprietary language models operating in interactive environments that integrate multiple data sources and external APIs through autonomous agents. We leverage Grouped Relative Policy Optimization (GRPO) to generate and evaluate multiple complete execution trajectories per query within a simulated interaction environment, assigning trajectory-level rewards through an autonomous LLM-based evaluation module (RULER). High-reward trajectories are used to iteratively update SLM parameters, producing a compact, specialized policy that improves tool selection accuracy, response quality, and robustness under dynamic query distributions. To address the limitations of fine-tuning proprietary models, we introduce a memory-based middleware that stores query-trajectory-reward tuples and reuses high-reward trajectories at inference time for semantically similar queries, enabling trajectory-level adaptation without modifying underlying model parameters. A learned reuse policy governs whether to replay, adapt, or regenerate trajectories, yielding substantial reductions in inference latency. Overall, the proposed architecture combines GRPO-based policy refinement with learned trajectory reuse, improving accuracy and efficiency while reducing reliance on expensive large-model retraining and preserving deployment flexibility across Agentic AI use cases such as database querying, web search, and multi-tool orchestration. We further validate the framework on a multi-agent information-seeking task involving enterprise email search (Enron corpus) with dynamic redirection to web search via SearXNG, demonstrating generalization beyond incident-management workflows.

      Agentic AI
      LLM
      Reinforcement Learning
      Read paper
      paper

      MedQwen-VGR1: Multimodal Vision-Guided Reasoning Model for Med-VQA, Temporal Diagnosis and Drug Interaction Analysis in Medical Imaging and Pathology

      ICECET 2026Subhadeep Mandal, Sukanya Bag, Ankit Sati, Chirag Jain

      MedQwen-VGR1 is a novel multimodal vision language model (VLM) trained for medical visual question answering (Med-VQA), longitudinal temporal diagnosis, and drug interaction analysis across radiology and pathology imagery. The model undergoes a multi-stage training pipeline comprising: (1) Continuous Domain-Adaptive Pretraining on heterogeneous medical image-text corpora; (2) Supervised FineTuning on expert-annotated multimodal datasets spanning sequential imaging conversations; (3) Human Preference Alignment via Direct Preference Optimization (DPO) and Group Relative Preference Optimization (GRPO) applied to Vision-Guided Chain-of-Thought (CoT) Reasoning trajectories; and (4) Smart Memory Module integration enables Cross-Visit Context Retention and Comparative Reasoning on sequential or temporal pathologies. MedQwen-VGR1 processes sequential medical images and textual queries to generate stepwise, evidence-grounded rationales correlating visual features across timepoints, predict adverse drug interactions from images/metadata, and output calibrated reliability scores. Evaluations on PathVQA, SLAKE, and VQA-RAD demonstrate superior performance over LLaVA-Med (+14.7% accuracy) and BioViL-T (+12.3% temporal reasoning), achieving 82.4% VQA accuracy and 78.6% change detection precision, enabling multi-turn clinical dialogues and automated clinical reports. The framework enables dynamic, multi-turn clinical dialogues with automated medical report synthesis, addressing critical gaps in temporal reasoning, explainability, and clinical safety alignment for real world AI and safety enhanced clinical decision support systems.

      Multimodal
      Medical AI
      Reinforcement Learning
      VLM
      Read paper
      paper

      Automating Insurance Claims Processing: Finetuning Tiny Language Model through QLoRA

      ICECET 2025Subhadeep Mandal

      This paper presents a low-cost Generative AI framework for automating insurance claims processing, transforming labor-intensive workflows into efficient, consistent operations that boost customer experience. Leveraging Phi-3.5 optimized via Quantized Low-Rank Adaptation (QLoRA) and a synthetic dataset from multiple LLMs, the system automates email-to-JSON conversion, entity extraction, and verdict generation (approvals/rejections) with transparent justifications. Trained on diverse writing styles for real-world robustness, it achieves 98% entity extraction accuracy and 99.63% verdict precision while adhering to Responsible AI principles for fairness and risk reduction. This scalable solution sets a new benchmark for transparent, high-performance claims automation in insurance.

      LLM
      QLoRA
      Read paper
      paper

      A Framework for Enhancing Product Recommender Systems with Knowledge Graph Embeddings, Multi Agents, LLM and Real-Time Image Search

      ICECET 2025Subhadeep Mandal, Sukanya Bag, Ayushman Gupta, Chirag Jain

      This paper introduces a novel multi-agent recommendation framework integrating Large Language Models, Knowledge Graph Embeddings, and real-time image search to overcome the limitations of standalone deep learning systems. Utilizing graph databases and vector stores for multimodal data management, the architecture deploys four specialized AI agents orchestrated by a central LLM. This enables direct interaction with external databases for advanced entity classification, link prediction, and real-time complex query processing. The resulting domain-agnostic system significantly advances the contextual accuracy and multimodal capabilities of personalized recommendation services.

      Agentic AI
      Knowledge Graph
      Recommendation System
      Computer Vision
      Read paper
      patent
      Filed

      Unified 7B Multimodal Reasoning Model for MedVQA, Temporal Diagnosis, Report Generation, and DDI Analysis

      USPTO · Application no. pendingSubhadeep Mandal, Sukanya Bag, Ankit Sati, Chirag Jain
      Multimodal
      Medical AI
      Reinforcement Learning
      patent
      Filed

      Speech2Text Agentic Framework for Custom Insights Generation from Audio

      USPTO · Application no. pendingSubhadeep Mandal, Sukanya Bag, Ankit Sati, Chirag Jain
      Multimodal
      Audio
      STT
      TTS
      patent
      Filed

      Agentic RL Framework to Fine-Tune SLMs and Improve Agentic Tool-Calling Accuracy

      USPTO · Application no. pendingSubhadeep Mandal, Sukanya Bag, Ayushman Gupta, Chirag Jain
      LLM
      Agentic AI
      Reinforcement Learning
      Contact

      Get in Touch

      Want to chat? Just shoot me a dm with a direct question on twitter and I'll respond whenever I can. I will ignore all soliciting.

      SMBusiness card
      Subhadeep Mandal
      Subhadeep Mandal
      Lead AI Engineer
      Axtria Ingenious Insights