Chandan Velji

Developed a bilingual, RAG-enabled financial assistant handling 50K+ monthly queries with real-time TASI market data. Engineered end-to-end ETL pipelines and APIs to ingest and normalize daily records, overcoming latency and data-relevancy challenges. Built a multimodal document analyzer agent for PDFs, enabling Q/A and summaries that cut review time by 50%. Fine-tuned LLaMA3 via QLoRA PEFT on a 30K-doc financial corpus, boosting domain-specific answer accuracy and reducing hallucinations.

Architected end-to-end AI/ML pipelines—covering data ingestion, feature engineering, model training, evaluation, and RESTful API serving. Designed and deployed a custom OCR pipeline for ancient Hebrew manuscript pages by training YOLO and VGG, achieving 98% character-level accuracy. Trained a quantized person detection and tracking model on a TPU-accelerated Raspberry Pi, delivering 15 FPS inference with 90% mAP for real-time edge robotics. Built a real-time forklift worker proximity alert system with YOLO.