Hi, I'm
Building AI systems that reason, retrieve & act
Minimal, purpose-built architectures spanning inference serving telemetry, agent tool protocol gateways, relational GraphRAG, and federated analytical flows.
Minimal, purpose-built systems spanning production serving telemetry, agent protocols, relational retrieval, and federated intelligence.
Load profiler for LLM endpoints — measures TTFT, throughput, and tail latencies on vLLM.
Production MCP server giving agents standardized, typed tool access across APIs and databases.
JSONSchema(strict=True)Hybrid retrieval engine combining Neo4j knowledge graphs with vector search for multi-hop queries.
Multi-agent system running ReWOO workflows and federated SQL via Trino — no data centralization.
The real reason your AI agent is slow isn't code — it's architecture. A deep dive into latency waterfalls, distributed search fan-outs, and parallel tool orchestration in production AI agents.
Exploring Anthropic's Model Context Protocol (MCP) as the universal standard for LLM tool invocation: strict JSON-Schema validation, client-server decoupling, and secure enterprise data access.
A guide to overcoming vector retrieval limitations: constructing entity-relation knowledge graphs with Neo4j and orchestrating multi-hop relational retrieval for accurate agent reasoning.
AI Engineer at Strategy specializing in LLMOps, compound AI systems, autonomous agent workflows (LangGraph, ReWOO, ReAct), GraphRAG, and production model serving with vLLM. Open to technical collaborations, engineering discussions, and high-impact roles.
curl -s https://cosmos127.dev/api/contact