Multimodal Contrastive Learning for Materials Science

CLIP and ImageBind adapted across crystal structures (CGCNN graph encoder), XRD diffraction patterns (1D ResNet), and natural-language descriptions (MatSciBERT) via InfoNCE loss in PyTorch + PyTorch Geometric. 94.0% Recall@1 on text-to-crystal retrieval, 90.5% Recall@1 on crystal-to-text, and MRR 0.966 on 9,751 Materials Project compounds. Cross-modal retrieval and zero-shot classification for materials discovery.

xarray + SQL + Lumen - Scientific Data Dashboards

206 tests · 10 transforms · 2 source classes · 2,300+ line dashboard

SQL queries on N-dimensional xarray datasets via Apache DataFusion. CF conventions auto-detection, AI analysis hooks, spatial binning at 68M cells in <5s.

ML Interpretability & Embedding Explorer

1,219 tests · 94% coverage · 8 interactive modules · 7,600+ LOC

SHAP explanations, counterfactual explorer, drift detection, error analysis. Built on Panel, HoloViews, Datashader, and Param.

Fashion AI + Virtual Try-On

HR-VITON with CBAM attention, CLIP embeddings + Pinecone search, LLaMA3 "StyleRAG" agent. SSIM 0.91, +18% CLIP similarity gain across 1,200+ test images.

AI-Powered Finance Platform

11 integrated modules spanning education, compliance, advisory, and automation. WealthWise (RAG on SEBI/BSE/Income Tax), Loan Recommender (LLM-ranked across 250+ Indian schemes), multi-agent Goal-Based Planner, FinEdu (Deep Knowledge Tracing), Mutual Fund Advisor, Scam Checker (Naive Bayes), Stock Alerts (Twilio WhatsApp), and Finnyland (gamified kids finance). Built on Flask, LangChain, LangGraph, Transformers.

Multimodal RAG for E-commerce

BLIP captions + OCR + ViLT reranking with Pinecone vector search. 97% response relevance, 30% fewer hallucinations, 22% better retrieval precision on 500+ product queries.

Speech & Language AI

Voice-first, multimodal AI platform with 3 specialized agents (Travel, Welfare, Mentorship). Multilingual, works on low-end phones without internet via calls, SMS, and WhatsApp.

MCP Server for HoloViz Visualizations

36 tools, 8 MCP Apps, Panel embed support, crossfiltering, and streaming - interactive visualizations via AI assistants.

Quantitative Trading

Sharpe Ratio 3.5+, 300%+ returns on BTC, 105%+ on ETH. 10+ statistical indicators and momentum strategies using Pandas, NumPy, TA-Lib, and Backtrader.