AI researcher and founder building memory, retrieval, and reasoning systems for AI agents — knowledge-graph retrieval-augmented generation (RAG), bi-temporal memory, and reinforcement learning for keeping agent memory current.
Founder & Chief AI Researcher at Vrin. Prev: SWE @ Microsoft (via Drevol), ML research @ UC Davis. Published at IEEE ISCAS.
-
Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents — Patel, V. (2026) An open reinforcement-learning environment that trains LLM agents to use the current fact, not the stale one. GRPO-tuned Qwen2.5-3B (LoRA) nearly doubled held-out accuracy: 9.0% → 16.7%. paper · code · Prime Intellect Hub
-
Extended Operational Life for Wearable Health Devices: A Hybrid TinyML and Server-Side ML Approach — Nazari, Patel, et al. · IEEE ISCAS 2025 IEEE Xplore
-
Engram ·
open source— cross-document reasoning for RAG ingestion: surfaces contradictions, derived facts, and implicit connections before chunks reach the embedder. Graph-aware retrieval planner + strategic per-query router. -
Vrin — multi-tenant RAG + agent platform on AWS (Neptune knowledge graph, OpenSearch, Lambda). Sub-second graph retrieval that beats the HippoRAG-2 SOTA on MuSiQue, with bi-temporal fact versioning and BYOK cross-account deployment.
Python · PyTorch · RL (GRPO / LoRA) · verifiers / prime-rl · vLLM · knowledge graphs · RAG · LangGraph · MCP · AWS
📫 vedantpatel.dev · X · LinkedIn