Gourab Roy crimsonKn1ght

My research focuses on building robust and interpretable architectures in data-scarce scenarios.
I'm particularly interested in explainable and trustworthy AI.

Projects

Vision-Language Models

PointCal-C: Low-cost selective zero-shot 3D recognition under corruption — audits how a frozen CLIP backbone's accuracy and confidence reliability degrade across corrupted point clouds, and fits a lightweight abstention rule on top.
EditCheck-VL: Low-cost calibrated counterfactual visual-change verification — a frozen SigLIP2/DINOv2 backbone plus a ≤1M-parameter probe decides whether a claimed before/after scene change holds, localizes it, and abstains when the evidence is ambiguous.
vlm-model: A minimal Vision-Language Model (frozen CLIP encoder + frozen LLM + trainable connector) following the LLaVA architecture.
medical-rag-vlm-ii: Year-1 evaluation harness for reducing hallucination in medical VLMs via retrieval-augmented grounding — runs pretrained medical VLMs on radiology VQA / report generation and scores outputs for factual consistency and hallucination.
medical-rag-vlm: Early-stage minimal LLaVA-style VLM scaffold for the medical RAG-VLM research line.

Astronomy / Space ML

AstraQ-VL: Minimal, efficient astronomy Vision-Language Model (LLaVA-style, CLIP + Qwen, LoRA instruction tuning) for solar/astronomical imagery.
TerraQ-VL: Sibling to AstraQ-VL — the same LLaVA-style VLM recipe adapted for Earth-observation / remote-sensing imagery (VRSBench).
astroclip-galaxy10: LLaVA-style VLM (CLIP + Qwen, LoRA instruction tuning) trained for astronomy image-text alignment on the Galaxy10 dataset.
galaxy-image-spectrum-clip: Cross-modal (image ↔ spectrum) representation learning on the Multimodal Universe (MMU) dataset.
space-net: Research-grade PyTorch pipeline for detecting, denoising, and localizing gravitational-wave chirp signals in noise, structured as a multi-level curriculum.

NLP / LLMs

Docqnatool: Docqnatool is a smart document Q&A assistant that lets you upload your files and ask meaningful questions about their content. Click for demo.
DocuMindGPT: DocuMindGPT is a document-grounded Q&A CLI that uses RAG (Retrieval-Augmented Generation) to answer questions based on your PDFs or text files. It features a built-in evaluation agent that scores answers for hallucinations and relevance.
rag-llm: Retrieval-Augmented Generation (RAG) demo built with LangChain, Chroma, and Streamlit, from scratch.
mini-gpt: Mini gpt based off Andrej's course.
llm-app-agent-frontend_v1: React + Vite frontend for an LLM chat agent with streaming responses, chat history, Markdown rendering, and optional web search.
llm-app-agent-data-science_v1: Async FastAPI service that routes questions through a LangGraph workflow and streams progress and answers as NDJSON, with conversation memory for follow-ups.

Vision

Knee osteoarthritis detection and categorization: Code for paper "Knee osteoarthritis detection and categorization with deep learning models". Click for paper.
Texture-based Feature Extraction and CBAM-Enhanced U-Net: Code for the paper titled "Texture-based Feature Extraction and CBAM-Enhanced U-Net for Automated Knee Osteoporosis Detection"
CT scan reconstruction: This repository implements an end-to-end CT image reconstruction pipeline that maps sinograms → reconstructed CT images using deep learning
Real-time object detector: A web-based real-time object detection application + standalone version which is a general purpose real-time detector
Image generation: Image generation using huggingface models.
Out of Sight, Not Out of Mind: Reproduction of OSNOM's Lift-Match-Keep method — tracks where objects are in 3D from egocentric video, including while they're outside the camera's view.

Diffusion / Generative

compact-clad: Reproduction of CLAD used as the starting point for a lighter formulation — tests whether the diffusion state can be made smaller.

Kaggle

ML Projects

Movie Recommender: A comprehensive movie recommendation system implementing multiple machine learning algorithms with an interactive React frontend.
Misc. AI / ML codes: Some AI/ML tools written by me.
model-pedia: Small, self-contained reference implementations of the models you meet on the way into deep learning — each one readable in a sitting and runnable in minutes on a laptop CPU.

Apps / Projects