About Code Metal

Code Metal is the leader in automated software engineering you can trust. As AI writes more of the world's code, the bottleneck in software has shifted from writing code to verifying it works, and AI cannot verify its own work with certainty. Code Metal takes a fundamentally different approach: constrain AI to what it does reliably, verify every step independently of the model using formal methods, and keep engineers in the loop on the decisions that matter. The result isn't code that probably works — it's code that is provably correct, with auditable proof. Customers including the U.S. Air Force, L3Harris, RTX, and Toshiba use Code Metal to modernize legacy code, optimize performance on real hardware, and move prototypes to production, fast. Founded in 2023 with offices in Boston and San Francisco, Code Metal is funded by Accel, Salesforce Ventures, B Capital, Smith Point Capital, J2 Ventures, Shield Capital, Overmatch, RTX, and others.

Learn more at codemetal.ai.

About the Role

We're building next-generation AI systems that help military planners explore, compare, and evaluate operational courses of action. Our work combines frontier language models, simulation, planning, and verification into human-in-the-loop decision-support systems for defense applications. As an Applied AI Research Engineer, you’ll focus on human machine teaming and agentic AI to build systems that allow warfighters, planners, analysts, and decision-makers to explore operational choices with speed, confidence, and control.

This role focuses on designing and building agentic AI systems – not chatbots. You'll develop multi-agent workflows, fine-tune and evaluate models, build retrieval pipelines, experiment with post-training techniques, and integrate AI with simulation and planning software. You'll work closely with AI researchers, software engineers, and defense experts to turn research ideas into production-ready capabilities. The goal is to make complex planning, wargaming, adjudication, and analysis workflows faster, more explainable, and more trustworthy.

Research Areas of Interest

An incomplete list of ongoing and near-term directions:

Human-machine teaming for AI-assisted course-of-action development, comparison, critique, refinement, and operational decision support

Agentic planning systems that integrate language models with simulation, doctrine retrieval, external tools, structured outputs, and deterministic verification

Adapting and optimizing foundation models through fine-tuning, post-training, distillation, reinforcement learning, and rigorous evaluation for planning and decision-support tasks

Multi-agent AI systems for Red/Blue planning, control-cell support, adjudication, branch-and-sequel analysis, and collaborative planning workflows

Building reliable AI systems using self-correction, structured reasoning, constraint-aware generation, verification, and robust tool use

Learning from human expertise through planner feedback, preferences, approvals, synthetic data generation, and human-in-the-loop improvement

Trustworthy AI for high-consequence applications, with an emphasis on explainability, provenance, traceability, auditability, uncertainty estimation, and model behavior analysis

What You’ll Do

Design and build agentic AI systems for planning, decision support, and human-machine teaming

Develop AI pipelines that integrate foundation models, retrieval, simulation, external tools, and deterministic software

Design, run, and analyze experiments to evaluate model and agent performance, reliability, traceability, latency, cost, and user trust

Fine-tune, distill, and evaluate foundation models for domain-specific planning, reasoning, and decision-support tasks

Build datasets, retrieval pipelines, automated benchmarks, and experiment infrastructure to support continuous model improvement and reproducible research

Partner with software engineers to transition research prototypes into scalable AI services

Collaborate with domain experts to translate operational workflows into AI-enabled capabilities while ensuring AI outputs remain explainable, reviewable, and under human control