Four end-to-end projects — retrieval-augmented QA, an MLOps pipeline with live drift monitoring, multimodal search, and explainable vision. Each is tested, containerised, CI-verified, and built to be cloned and run.
Ask any question about the Constitution of India; get a grounded answer that cites the exact articles it came from. Article-aware chunking, optional cross-encoder re-ranking, and an evaluation harness (100% Hit@5 on the bundled corpus).
Predicts IPL match outcomes and monitors itself for data drift. Three containerised services (FastAPI + PSI drift monitor + live dashboard) share state through a volume, with a real serving-aligned drift reference and quantile-binned PSI.
Find any Indian dish by describing it in plain English or uploading a photo. Zero-shot CLIP embeddings mapped into a shared text–image space, with cosine retrieval and a category breakdown of every result.
Upload a leaf photo, get a disease diagnosis across 15 crop-disease classes — with a Grad-CAM heatmap that shows exactly which region of the leaf drove the prediction, so the model's reasoning is auditable.