Repository containing end-to-end machine learning projects on object detection (computer vision) and local retrieval-augmented generation (RAG) systems, including their evaluation and deployment.
RT-DETRv2 fine-tuned to detect trash, hands, and bins in real-world photos, with three deliberate hard-negative classes (not_trash, not_hand, not_bin) so the model learns to reject near-misses rather than just recognize positives. Deployed as a gamified Gradio demo.
- Live demo: Trash Object Detection
- Result: mAP@50 of 0.524 on a held-out test split; weaker on medium-sized objects (mAP 0.13), a known limitation tied to dataset scale.
A RAG pipeline built without any retrieval framework — chunking, embedding, semantic search, and generation implemented directly with PyTorch and HuggingFace primitives, using Gemma as the local LLM. A reference notebook, not a deployed demo — the deliberate "from scratch" companion to the framework-based systems below.
Answers questions about uploaded PDFs by retrieving from both the document's text and captioned descriptions of its images, using a 4-bit quantized google/gemma-3-4b-it for both captioning and grounded answer generation, backed by a FAISS index.
- Live demo: Multimodal RAG System
Each project folder has its own README covering pipeline design, model/training details, evaluation results, engineering decisions, and known limitations.
├── object_detection_CV/
│ ├── Drawing_Bounding_Box.ipynb
│ ├── Object_Detection_Notebook.ipynb
│ └── README.md
├── rag_systems/
│ ├── local_rag/
│ │ ├── Local_RAG_from_Scratch.ipynb
│ │ └── README.md
│ ├── multimodal_rag/
│ │ ├── Local_Multimodal_RAG.ipynb
│ │ ├── MRAG_HF_Upload.ipynb
│ │ └── README.md
│ └── README.md
├── README.md
├── .gitignore
└── LICENSE
MIT — see LICENSE.