- Streamlit-based web RAG chatbot (LangChain + Pinecone, text-embedding-3-large)
- Added reranking (Cohere) and improved Precision 0.40 → 0.70 on 100 QA pairs
- Guardrails via few-shot prompting + query preprocessing (dictionary/normalization)
- Fine-tuned DocLayout-YOLO + reading-order reconstruction
- W&B-tracked sweeps for category-wise NMS IoU thresholds → public score +0.051 (0.371 → 0.422)
- Layout detection improved by +0.475 F1 on multi-column documents
- Implemented single-pass generation → self-evaluation → selection prompting
- ~1.5× faster inference than ToT under an SLM setup (CommonsenseQA/OpenBookQA)
