Hi, I'm Prajwal 👋

Built a fully client-side Vision-Language Model (VLM) system using Qwen 3.5 (0.8B/2B), enabling OCR and multimodal reasoning directly in the browser via ONNX Runtime and WebAssembly — eliminating all backend inference dependency. Optimized performance through Q4 quantization, balancing model size, latency, and accuracy under browser memory constraints.

Integrated the Unstructured API into the QiLegal enterprise backend to support end-to-end document ingestion and preprocessing pipelines (PDF, DOCX, emails). Designed granular logging systems across the codebase to improve observability, debugging, and production reliability.

Worked with Vespa vector database for semantic search and retrieval in QiSearch — querying indexed legal datasets, analyzing centroid-based clustering, and supporting approximate nearest-neighbor retrieval for case similarity and legal judgment search.

Developed a production-grade web scraping system for legal data extraction, featuring multi-page crawling, dynamic content handling, structured data parsing, retry/failure mechanisms, and clean data pipelines for downstream AI indexing and RAG ingestion.

Contributed to a production-level RAG (Retrieval-Augmented Generation) pipeline integrating document processing, embedding generation, Vespa vector storage, and LLM-based response synthesis for intelligent legal query handling across QiLegal's platform.