Aryan Kekre

Document Verifier SLM (Deep Learning / Transformer Architecture)

Built a production-style identity document verification system using a custom Dual-Encoder Cross-Attention Transformer trained from scratch in PyTorch on 500K synthetic samples. Designed a character-aware tokenizer and transformer architecture achieving 93% test accuracy for Aadhaar/document forgery detection. Implemented curriculum learning, EMA, focal loss, and hard negative mining; developed a three-tier inference pipeline with explainable field-level attribution.

View details

Neural Archive v3.0 (RAG & LLM Infrastructure)

Engineered an end-to-end RAG platform using FastAPI, ChromaDB, Gemini embeddings, and IBM Docling for OCR-aware ingestion of technical documents. Implemented semantic chunking, vector persistence, neighboring-context retrieval, and 3D embedding visualization via PCA + Plotly.

View details