AI Engineer, Agents and Evaluation

Pronoia•Torrance, CA

7d•Onsite

About The Position

Pronoia Energy is seeking an AI Engineer to build an internal, grounded, agentic AI system that accelerates research. This is a founding, high-leverage role where you will own the part of the stack that decides what counts as true before the team acts on it. You will build the agent orchestrator that plans a research question into steps and routes each step to the right tool, and the first verification gate and evaluation harness that catch the system before it states anything unsupported. The core principle is: factual questions go to retrieval, computational questions go to tools, and nothing is answered from the model's own memory. You will stand up a self-hosted open-weights model on our own hardware and build the entire agent loop around it. This is a founding generalist role, and early on you will wear several hats across orchestration, serving, and evaluation. This role is onsite in Torrance, California.

Requirements

A track record shipping a real agent or tool-use system in production, not a demo.
Genuine experience self-hosting open-weights models on your own GPUs, not only calling a hosted API.
A habit of treating hallucination and tool-misuse as engineering problems with mechanical fixes.
Builder-generalist range across orchestration, model serving, and evaluation.
A self-starter who thrives amid ambiguity and rapid iteration.

Nice To Haves

A scientific or physics background that makes our zero-fabrication standard second nature.
Contributions to open-source serving or agent tooling such as vLLM, TGI, or agent frameworks.
Experience building retrieval-augmented systems and evaluation suites.
Prior work in deep tech R&D alongside interdisciplinary teams.

Responsibilities

Build the agent orchestrator and the tool-routing logic at the heart of the system.
Stand up and serve a self-hosted open-weights model on local hardware.
Build the first verification gate and evaluation harness, making grounding mechanical rather than hopeful.
Enforce strict routing so claims trace to a retrieved source or a deterministic computation.
Wear several hats as a founding generalist and later hand specialties to dedicated hires.