Kushal Desai

Architected JoyGPT, a production-grade multi-tenant AI chat platform on FastAPI and LangGraph for ~100 transport operators. Implemented context-aware LangGraph routing, an end-to-end RAG pipeline indexing 130+ operator documents and 29 rider documents per tenant, and token accounting with MongoDB-backed JWT RBAC plus Redis-backed distributed locking for async sync jobs.

Built an autonomous LinkedIn Easy Apply/Long Apply agent that processed ~50 applications per run, cutting recruiter effort from half a day to a single automated execution and delivering 500%+ throughput. Implemented LangChain + Groq-based context-aware form answering and a RAG pipeline to score and rank candidate profiles against job descriptions for shortlisting.

Developed custom YOLO and CNN models for real-time object detection in live sports broadcasts, achieving a 15% accuracy improvement over an 86% baseline and reducing false positives by 30%. Integrated the models into the production broadcast analysis pipeline in collaboration with the core engineering team.

Built a multi-agent autonomous web navigation system that interprets natural-language queries, selects target sites, and returns structured answers. Orchestrated the workflow with LangChain using LLaMA 3.2 Vision and OpenAI models, with Microsoft OmniParser extracting interactive page elements for LLM-driven action selection and Selenium executing those actions.

Developed a YOLOv8-powered traffic violation detection system identifying helmet absence, triple riding, and lane violations in real time, using NanoNets OCR for license plate recognition and automated penalty email notifications. Improved enforcement efficiency by 40%.

Earned a Bachelor of Engineering in Computer Science from Indus Institute of Technology & Engineering.