Antoine Zambelli is an AI engineering leader and the creator of Forge, a reliability layer for tool-calling language models. His work focuses on the failure points between a model choosing a tool and an application completing a multi-step task.[2][3]
Start with Forge →
Career
Zambelli’s self-maintained ORCID record connects financial engineering with production AI, distributed inference and engineering leadership. His career includes quantitative work, a company he founded, and data-science leadership before his current AI Solutions role at Texas Instruments.[3]
Contributions
Repair failures inside a tool-calling loop
Forge validates model responses, attempts to recover malformed tool calls and supplies retry guidance. Optional prerequisites and required steps constrain a workflow when the application needs them. It can run as middleware, a workflow runner or a proxy between an existing client and a model backend. The developer still supplies the inference backend and tools. Its published repository results concern its own scenario suite and particular model configurations; they do not establish reliability for every application.[2]
Make an external role execution inspectable
His Codex role runner launches named roles in separate CLI sessions and retains a handle, requested configuration and returned results. A caller can inspect, wait for or continue a run. The design makes the task and working directory explicit because the external session does not inherit a parent conversation’s hidden context. It is an unofficial developer tool for a configured local CLI environment.[4]
Tools, writing & talks
Tools
Start here
Forge
github.comA Python framework and standalone proxy adding validation, recovery and workflow constraints to LLM tool calls.
Start with the proxy-versus-workflow-runner comparison, then test one small tool loop with a known expected result. The Python library needs Python 3.12 or later; all installation modes still need a configured model backend and suitable access.
Creator, identified in repository copyright and publication credit[2]Original source- codex-exec-subagentsgithub.com
An external role runner with persistent handles and retained execution evidence.
Repository author[4]