Shounak Das

Final year Dual Degree (B.Tech + M.Tech) student in Electrical Engineering at IIT Bombay, with a minor in AI, Machine Learning & Data Science, graduating in 2026.

I am currently working on Vision-Language Models (VLMs) at MeDAL Lab under the guidance of Prof. Amit Sethi. I have worked on cutting-edge industrial problems in ML, NLP, Generative AI, LLMs, and Computer Vision at Fujitsu Research, Intel, and Swiggy. I am also proficient in DSA and programming, and skilled in Signal Processing and Communication Systems.

Skills:

• Languages: Python, C++, C, MATLAB, SQL, JavaScript, HTML, CSS, Bash
• ML & Deep Learning Frameworks: PyTorch, TensorFlow, scikit-learn, Keras, Hugging Face, Diffusers, LangChain, LangGraph, OpenVINO
• NLP & Computer Vision: spaCy, NLTK, Gensim, OpenCV, Whisper, CLIP, Gemini API, OpenAI API, gTTS
• Data & MLOps: PySpark, ChromaDB, Weaviate, Docker, AWS, Databricks, Linux, Git, MLflow, Snowflake
• Deployment & Web: Django, React, Streamlit, REST APIs, FastAPI, Flask, Nginx, Postman

I have a strong interest in Generative AI, NLP, Computer Vision, and Machine Learning and enjoy working on projects in these areas. I’m excited to apply my skills to impactful projects.

Publications:

• ICML '25 (International Conference on Machine Learning)
• MIDL '25 (Medical Imaging with Deep Learning)
• ISBI '25 (International Symposium on Biomedical Imaging)
• ICPR '24 (International Conference on Pattern Recognition)
• BIBE '24 (International Conference on Bioinformatics & Bioengineering)
• BioImaging '24 (International Conference on BioImaging) — Best Student Paper Award

Education:

• Dual Degree (B.Tech + M.Tech) in Electrical Engineering with a minor in AI, ML & Data Science at IIT Bombay

Email  /  Google Scholar  /  GitHub  /  LinkedIn  /  X

• [06/2025] Our paper FEDTAIL has been accepted to ICML 2025!
• [05/2025] Our paper Whole Slide Image Domain Adaptation Tailored with Fisher Vector, Self Supervised Learning, and Novel Loss Function has been accepted to MIDL 2025!
• [01/2025] Our paper Scalable Whole Slide Image Representation Using K-Means Clustering and Fisher Vector Aggregation has been accepted to ISBI 2025!
• [09/2024] Our paper Clustered Patch Embeddings for Permutation-Invariant Classification of Whole Slide Images has been accepted to BIBE 2024!
• [06/2024] Our paper IDAL: Improved Domain Adaptive Learning for Natural Images Dataset has been accepted to ICPR 2024!
• [02/2024] Best Student Paper Award for our BioImaging 2024 paper on domain adaptation for histology images!
• [12/2023] Our paper Domain-Adaptive Learning: Unsupervised Adaptation for Histology Images has been accepted to BioImaging 2024!

• Developed a proactive RCA framework for ~100k-line Warrior logs for next-step prediction and failure diagnosis.
• Built an InstructRAG pipeline using Gemini to synthesize QA data and fine-tuned Mamba-2 (2.7B) with LoRA.
• Overcame LLM context limitations using hypergraphs with Personalized PageRank and similarity compression.
• Achieved 42% BLEU and 4.4/5 LLM-as-a-Judge score, and deployed into Fujitsu's Kozuchi platform for business use.

• Engineered an OCR system for Hindi, Telugu & Kannada using ViT to extract robust image features, enabling accurate recognition across diverse regional scripts.
• Integrated IndicBERT for context-aware recognition, boosting regional text digitization for NIC (Govt of India) and improving automated document processing.
• Utilized WhisperX for multilingual audio labeling, enabling accurate transcription alignment and timestamp generation.
• Optimized inference speed by 67% (LLaMA-3 8B) and 80% (Mistral 7B) using OpenVINO and IPEX, improving overall AI workload performance.

• Developed a robust topic modeling framework using BERTopic and an Azure OpenAI LLM agent, enabling the identification & prediction of emerging trends in events and items, significantly enhancing predictive analytics.
• Built a custom spell-error dataset from Instamart SQL database using edit distance, decompounding, and phonetics.
• Developed a spell correction pipeline leveraging unigram and bigram probability models with fuzzy logic, achieving up to 83% correction accuracy.

• Designed and tuned ChromaDB and Weaviate pipelines for RAG, improving retrieval quality via optimized embeddings, indexing, and query-time filtering.
• Integrated TraceLoop and IBM Instana for VectorDB observability and real-time performance analytics.

• Worked on the MLCommons AlgoPerf Training Benchmark, a global competition focused on ML algorithms.
• Optimized a novel training algorithm across 6 datasets: Criteo 1TB (clickthrough), FastMRI (reconstruction), ImageNet (classification), LibriSpeech (speech), OGBG (molecular), and WMT (translation).

• Implemented models like Deforum Stable Diffusion and VQGAN for text-to-image and image-to-image tasks, generating high-quality visual content from prompts.
• Developed image-to-animation pipelines and deployed a prompt-based API on AWS for real-time inference, enabling interactive generative workflows.
• Explored advanced flow-based models like RealNVP to enhance generative content diversity and quality.

• Developing a multi-resolution vision-language pipeline for gigapixel images, enabling scalable text-guided representation learning.
• Engineering semantic-guided prompt tuning with CLIP and LLaVA, using knowledge distillation for robust few-shot classification.
• Implementing distribution-aware cross-modal alignment to reduce modality gaps and improve generalization in VLMs.

• Developed an AI Guard Agent for real-time surveillance using vision, speech, and LLM-based reasoning.
• Integrated Whisper, Coqui TTS, and Gemini with meta-prompting and 4-level escalation for access control.
• Achieved 97.6% SSIM for face image similarity verification and 0.8s speech-to-text latency with robust multimodal response.

• Developed a WSI classification framework using attention-based MIL on patch-wise Fisher vectors.
• Extracted patch features via ResNet, Swin, MoCo, SimCLR and encoded using a 5-component GMM.
• Achieved AUC 0.83 (Warwick) and accuracies 0.86 / 0.84 on TCGA-BRCA / LUAD, surpassing SoTA benchmarks.

• Secured 1st place in the course image super-resolution Kaggle competition using a custom EDSR model on gaming data.
• Trained EDSR from scratch using ResBlocks and PixelShuffle, achieving a top score of 59.39 (joint PSNR + SSIM).
• Built vision models with CAM for CNN interpretability and generative models (GANs, Hierarchical VAEs).

• Built an NLP-based summarization system using TF-IDF, Seq2Seq, and BART with real-time deployment through Streamlit.
• Achieved ROUGE-L scores of 0.878 (BART) and 0.814 (Seq2Seq) on the XSum and SamSum datasets.

• Developed a Deep RL framework combining DQN, PPO, and FinBERT-based NLP sentiments for dynamic portfolio optimization.
• Engineered a custom Gymnasium trading environment using volatility, MACD, and RSI for risk-adjusted optimization.
• Achieved 25% annualized returns in backtesting on historical S&P 500 (SPY) data sourced from Yahoo Finance.

• Built a full-stack Inventory Management System using Django + React with secure authentication and multi-scanner integration.
• Designed and implemented REST APIs for real-time synchronization, concurrent data logging, and automated email alerts.
• Implemented a battery management module with LCD UI using C and the Raspberry Pi SDK for firmware-level control.

We present FedTAIL, a federated domain generalization framework designed to tackle domain shifts and long-tailed class distributions. By aligning gradients across objectives and dynamically reweighting underrepresented classes using sharpness-aware optimization, our method achieves state-of-the-art performance under label imbalance. FedTAIL enables scalable and robust generalization in both centralized and federated settings.

We introduce a domain adaptation framework for Whole Slide Image (WSI) classification that combines self-supervised learning, clustering, and Fisher Vector encoding. By extracting MoCoV3-based patch features and aggregating them via Gaussian mixture models, our method forms robust slide-level representations. Adversarial training with a hybrid PLMMD-MCC loss enables effective domain alignment, achieving strong performance on cross-domain HER2 classification tasks, even under label noise.

We propose a scalable method for whole slide image (WSI) classification that combines patch-based deep feature extraction, clustering, and Fisher Vector encoding. By modeling clustered patch embeddings with Gaussian mixture models, our approach generates compact yet expressive slide-level representations. This enables robust and accurate WSI classification while efficiently capturing both local and global tissue structures.

We propose a novel unsupervised domain adaptation (UDA) approach for natural images that combines ResNet with a feature pyramid network to capture both content and style features. A carefully designed loss function enhances alignment across domains with multi-modal distributions, improving robustness to scale, noise, and style shifts. Our method achieves superior performance on benchmarks like Office-Home, Office-31, and VisDA-2017, while maintaining competitive results on DomainNet.

We propose an efficient WSI analysis framework that leverages diverse encoders and a specialized classification model to produce robust, permutation-invariant slide representations. By distilling a gigapixel WSI into a single informative vector, our method significantly improves computational efficiency without sacrificing diagnostic accuracy. This scalable approach enables effective utilization of WSIs in digital pathology and medical research.

We propose a novel approach for unsupervised domain adaptation designed for medical images like H&E-stained histology and retinal fundus scans. By leveraging texture-specific features such as tissue structure and cell morphology, DAL improves domain alignment using a custom loss function that enhances both accuracy and training efficiency. Our method outperforms ViT and CNN-based baselines on FHIST and retina datasets, demonstrating strong generalization and robustness.

Website template borrowed from here.