Federico Kamelhar

I design and build production AI platforms: agent runtimes, multi-agent orchestration, and the infrastructure underneath them. Senior Principal Architect for Agentic AI at Oracle; I led the creation of Strands Agents at AWS, and I am building Tulip Agents, an open-source agentic framework. Based in New York.

I write here about infrastructure I run myself, and what measuring it showed.

What I work on

Each of these says what it is, what my part in it was, and where you can check.

Multi-agent orchestration. Governed runtimes where an orchestrator graph coordinates specialist agents under playbooks, tool-call admission gates and grounding checks — systems where the model proposes and the runtime decides. This is my day work at Oracle and most of it is not public; what is public is the paper below and the upstream contributions.

Grounding and hallucination control. Sole author of GSAR: Typed Grounding for Hallucination Detection and Recovery in Multi-Agent LLMs. The paper is the inspectable artefact: its method, its evaluation and its numbers are what I can stand behind publicly. An implementation of the design runs in production on internal workloads, which I can describe but not show you.

Agent frameworks. I led the creation of Strands Agents at AWS, now open source. I am currently building Tulip Agents, an open-core agentic framework.

LLM infrastructure, upstream. Oracle Cloud support in other people's projects, which is where the work is easiest to verify — merged pull requests, under my name:

• LiteLLM — the OCI provider, and fixes to its Cohere tool-calling path. 5 merged pull requests.
• oracle/langchain-oracle — core maintainer. 67 merged pull requests.
• vLLM production-stack — OKE deployment support, end to end. 2 merged pull requests.
• NeMo Agent Toolkit — OCI LangChain support for hosted Nemotron workflows. 2 merged pull requests.

Also author of Oracle's tutorial on deploying the vLLM production stack on OKE.

Elsewhere

Views here are my own and do not represent my employer.