CV

This is a description of the page. You can modify it in '_pages/cv.md'. You can also change or remove the top pdf download button.

Contact Information

Name Suparnojit Sarkar
Professional Title Applied Scientist 1
Email suparnojit2026@gmail.com
Location , Bengaluru, Karnataka

Professional Summary

Applied Scientist 1 at Glance, working on LLMs, computer vision, and generative AI.

Experience

  • 2026 -

    Bengaluru, India

    Applied Scientist 1
    Glance AI (InMobi Group)
    • Worked on conversion of the Z Image base model from T2I to I2I using training-free techniques.
    • Worked on realism for image generation using LoRAs applied on Flux Klein and Qwen Image.
    • Working on a Multi Tryoff-Flatlay-VTON PoC along with moderation handling at inference time.
  • 2026 -

    Remote

    Research Collaborator
    Cohere Labs
    Part of the Expedition Aya project.
    • Working on Tiny Aya Vision and safety representations across languages on Tiny Aya models.
    • Implemented a cross-modal weight merging strategy interpolating multimodal and base LLM weights with a tunable alpha, recovering text-only capabilities while preserving visual grounding performance.
    • Calculated refusal rate across all languages on Tiny Aya variants using LLM-as-a-judge (Command A) on the X-Safety benchmark.
  • 2025 - 2025

    Bengaluru, India

    Project (Gen AI/ML) Intern
    CINT Lab, Indian Institute of Science
    • Implemented end-to-end pipelines for image/video generation using Flux.1-Dev, Qwen Image, and Wan 2.2 (text-to-image and image-to-video).
    • Fine-tuned Flux.1-Dev (LoRA) for Studio Ghibli-style image generation with 80% accuracy.
    • Applied zero-shot, few-shot, and CoT prompt engineering for image/video generation on ComfyUI.
  • 2024 - 2025

    Mandi, India

    ML/CV Intern
    VIML Lab, Indian Institute of Technology Mandi
    • Worked on semantic segmentation of unmanned aerial vehicle remote sensing images under the guidance of Prof. Dinesh Singh.
    • Trained and evaluated the UAVid dataset using SegFormer and UNetFormer for image segmentation with PyTorch-Lightning, achieving an F1 score of 0.65.

Education

  • 2022 - 2026

    Kolkata, India

    Bachelor of Technology
    Heritage Institute of Technology
    Computer Science

Publications

  • 2026
    MedAidDialog: A Multilingual Multi-Turn Medical Dialogue Dataset for Accessible Healthcare
    MeLLM Workshop, ACL 2026
  • 2026
    IndicMedDialog: A Parallel Multi-Turn Medical Dialogue Dataset for Accessible Healthcare in Indic Languages
    BioNLP Workshop, ACL 2026
  • IndiaCultureVL-Bench: Cultural Recognition and Prompt-Induced Collapse in Open Vision-Language Models
    In review at a workshop, EMNLP 2026

Projects

  • Fashion-Search

    Built hybrid fashion retrieval using CLIP embeddings, K-means color extraction, and zero-shot scene classification with weighted re-ranking, enhanced with hierarchical indexing and inverted indices for 80% candidate filtering. Designed a semantic caching layer achieving 75-85% cache hit rates for sub-20ms response times. Fine-tuned CLIP ViT-L/14 with contrastive learning and hard negative mining, improving fashion-specific query precision by 15-20%.

    • Python
    • PyTorch
    • Hugging Face Transformers
    • ChromaDB
    • VLM
  • LegalAI

    Developed a multi-agent legal document analysis system with 4 specialized AI agents (LegalAdvisor, ContractAnalyst, LegalStrategist, Team Lead) that process PDF contracts to deliver structured risk assessments, compliance checks, and actionable insights via a Streamlit UI. Integrated ChromaDB with Gemini embeddings for semantic search, configurable chunking, dual-party contract analysis, risk-based clause highlighting, and real-time chat.

    • Python
    • Agno AI
    • Streamlit
    • ChromaDB
  • Polygon Colorizer

    Built a diffusion-based conditioned U-Net (UNet2DConditionModel) to colorize grayscale polygon images using CLIP text embeddings from color descriptions, using PyTorch, Diffusers, and Weights & Biases for training, experiment tracking, and visualization.

    • PyTorch
    • Diffusers
    • Weights & Biases
  • SLM for English Stories

    Built a 30M-parameter Small Language Model (SLM) for short English story generation, designing the pre-training pipeline from data preprocessing to autoregressive text generation, and evaluating a transformer-based text generation workflow.

    • PyTorch
    • Transformers
    • SLM Pre-training

Awards

  •
    HackNITR 5.0
    DSC NIT Rourkela

    Secured first place out of 3000 participants at one of Eastern India’s largest hackathons.

  • HackNITR 7.0
    DSC NIT Rourkela

    Mentored and judged AI/ML projects at one of Eastern India’s largest hackathons.

  • Hack Heritage 2023
    Heritage Institute of Technology, Kolkata

    Top 30 finish.

  • Amazon ML Challenge

    Top 2.5% (ranked 2013) among 82,796 participants.

Skills

Languages & Frameworks: Python, SQL, PyTorch, FastAPI, Agno AI, Crew AI
Tech & Dev Tools: Git, Docker, Dune, HuggingFace, ComfyUI, AI Toolkit, Unsloth, Streamlit, ChromaDB, ArgoCD
Libraries & Techniques: Fine-tuning, Prompt Engineering, Quantisation, Segmentation, RAG, LLMs, Agentic AI