Projects

RAG System

• Designed and built a complete retrieval-augmented generation pipeline for PDF, DOCX, and text documents: ingestion, preprocessing, cleaning, hierarchical parent-child chunking, embedding generation, and indexing.
• Implemented hybrid retrieval (vector similarity + keyword search) and hierarchical context reconstruction for grounded response generation.
• Built RBAC-aware retrieval with metadata-level access control so users only retrieve authorized documents/chunks across hierarchy levels.
• Implemented vector storage in Weaviate and parent-context persistence in PostgreSQL for scalable retrieval and reconstruction.
• Developed automated evaluation using RAGAS with faithfulness, answer relevancy, context precision, and context recall metrics.
• Shipped a GUI-driven user workflow with FastAPI backend for document upload and conversational querying.

Constrained Payment Gateway Routing

Improved payment success rate by 84 bps absolute under business constraints, enabling zero downtime during sale events and contributing $1.1M annual savings plus $30M incremental revenue.

Credit Risk Platform for 1P EMI

Batch and realtime risk models using graph embeddings, sequential behavior, domain adaptation, Flipkart activity, bureau signals, and SHAP explainability; reduced credit loss by 11% relative and increased whitelist by 41%.

Realtime Card Scanning AI

Built OCR-alternative system using synthetic data, EfficientNet + YOLO, Android integration, FastAPI on Azure, and model-size distillation from 245 MB to 1.2 MB.

Reinforcement Learning for Click through Rate Lift

Developed deep reinforcement learning model at Lemnisk for Ramanujan, improving CTR by around 30% and distilling behavior into interpretable decision trees for stakeholders.

Representation Learning and Feature Compression

Reduced 3000 features to 128-dimensional embeddings via transformer autoencoder while preserving model quality, and built SMS sender embeddings that reduced risk-model latency from 500 ms to 10 ms.