AI case studies: chest X-ray opacity screening, medical vision-language modelling, Tifinagh OCR at >99% accuracy, RAG chatbots, and drone simulation.
A sophisticated multi-modal AI system combining Vision Transformers and RoBERTa through co-attention. Enables fusion of radiology images and clinical text for better diagnostics. Pre-trained on ROCO (81k image-text pairs) and fine-tuned for VQA-RAD.
Stack: Vision Transformers, RoBERTa, PyTorch, Medical Imaging, Masked Autoencoding. Status: In Production.
A high-fidelity simulation of an autonomous delivery drone using ROS2, Gazebo, and YOLO. Includes flight control, object detection, obstacle avoidance, and DDS communication.
Stack: ROS2, Gazebo, YOLO, OpenCV, Python, C++. Status: Completed.
A full OCR pipeline for Tifinagh script built with CNNs, PaddleOCR, OpenCV preprocessing, and Flask + React frontends. Dataset includes 25,740+ images with various writing styles.
Stack: CNN, PaddleOCR, TensorFlow, OpenCV, Flask, React. Status: Completed.
An intelligent RAG chatbot integrating HuggingFace LLMs (Flan-T5, BART), LangChain, ChromaDB, and a React chat UI. Supports document uploads, summarization, semantic search, and real-time conversation memory.
Stack: LangChain, ChromaDB, Flask, React, HuggingFace LLMs, LLM Mistral AI. Status: Completed.
A technical workstation designed to bypass cloud latency and privacy concerns. It integrates deep PDF layout recognition with a local RAG pipeline and a pre-cached audio narration system. Specifically engineered to manage VRAM-heavy tasks like concurrent LLM inference, document parsing, and speech synthesis on an NVIDIA RTX 3050.
Stack: FastAPI, Next.js, Llama-3 (Ollama), Marker-PDF, Kokoro TTS, CUDA, ChromaDB. Status: Live / Local Deployment.
A two-stage AI pipeline for chest X-ray opacity screening, currently in a testing phase with physicians — not yet a certified medical device. A ResNet50 classifier first flags suspicious opacity, then a YOLOv8 detector localizes it, and a physician reviews every result before a report is generated. The pipeline is exposed two ways: a Streamlit wizard (patient intake → AI analysis → results & report) and a FastAPI endpoint for LAN/cross-machine access, both sharing identical preprocessing and inference logic. DICOM inputs go through VOI LUT and MONOCHROME1 inversion before CLAHE contrast enhancement, matching the exact pipeline used during training. Doctor feedback is logged to CSV for future retraining, and a PDF report is generated for each case via reportlab.
Stack: PyTorch, YOLOv8, FastAPI, Streamlit, OpenCV, pydicom. Status: In Testing — Physician Review.
An automated short-form video channel where a seven-agent review layer decides what gets made, what gets rewritten and what never ships. Open-source generators turn a topic into an MP4 in a single pass, with nothing checking whether the hook is worth watching, whether the footage matches the narration, or whether the rendered file even has an audio track — survivable when a person reviews each video, not when a channel publishes unattended every morning. I extended an MIT-licensed generator (renderer, web interface and provider integrations retained under its licence) with an original twelve-stage pipeline: agents screen scripts for policy risk before any render time is spent, score and rewrite weak hooks, query the real stock library for every search term, and inspect the finished file before it publishes. Around 2,900 lines of Python with 100 tests over the new layer.
Stack: Python 3.11, Google Gemini, Streamlit, FastAPI, MoviePy, YouTube Data API, Redis, pytest. Status: In Production.