Vaibhav Patel

Built production-grade LLM systems for the ADIA management board. Developed multi-agent architectures integrating retrieval, planning, and specialized task agents to support high-stakes decision workflows. Implemented retrieval, search/insight, and orchestration agents using Semantic Kernel, Agent Development Kit, and MCPs, with focus on grounded enterprise knowledge access, multi-source context orchestration, and real-time workflow execution. Deployed in live board workflows in a real organizational environment. Explored small language model (SLM) pretraining and end-to-end LLM implementation in PyTorch to build first-principles understanding of transformer training and inference.

• ◦ Worksmart: Designed and deployed memory and retrieval pipelines on AWS for extracting and retrieving contextual signals from screenshots, cursor activity, GitHub usage, and social data. Built agentic systems using FastMCP and LangGraph for retrieving "AI usage memories" at worker and team levels. Applied RAG and vector databases for semantic memory retrieval, trend discovery, and cross-session contextualization. Work involved practical challenges in retrieval quality, semantic matching, contextual memory, and knowledge access over noisy activity data.

• ◦ ALP Tempo: Led design and deployment of a secure, scalable microservices ecosystem using Node.js, TypeScript, AWS Lambda, Step Functions, and infrastructure-as-code. Improved system efficiency by 40% through better architecture and delivery workflows.

• ◦ Devflows: Built product features in Python-Django to support software development lifecycle workflows.

• ◦ Rax Key-Insights: Developed an AI-powered summarizer for research papers, employing traditional NLP techniques (tokenization, POS tagging, dependency parsing) to extract key insights.