Things I've built outside of papers, mostly to learn how something works end to end. Code is on GitHub.
Ask questions of a pile of research papers
A retrieval-augmented generation pipeline built from scratch. It splits papers along their real structure instead of fixed token windows, searches with FAISS, and answers with Llama 3.1 through Groq.
Find hairline cracks with a text prompt
Fine-tuned CLIPSeg to find cracks and taping joints from a text prompt, across ~6,000 images. Cracks are under 1% of pixels, so a Focal + Dice loss did most of the work (mIoU 0.52, Dice 0.67).
From PyTorch model to TensorRT engine
ResNet50 vs. Vision Transformer on 5,335 eye images (84.2% accuracy), then the whole road to production: PyTorch → ONNX → TensorRT, packaged in Docker.
Type a sentence, get a picture
A small Streamlit app on top of Stable Diffusion XL through the Hugging Face Inference API.
Real-time detection on small hardware
MobileNetV2 fine-tuned for edge devices, then quantized and optimized with TensorRT for live video.
Counting trees from the sky at ISRO
Found tree tops in rasterized LiDAR layers and grew regions out from them to outline each crown in a dense forest.