Location: San Francisco / Palo Alto (5 days a week in-office)

🚀 About Echelon AI

We’re automating the $1T IT services industry by building digital labor that autonomously builds, tests, and deploys enterprise software.

The difference: While most AI companies sell productivity tools, we’re building RL-driven, self-correcting AI agents that plan, implement, validate, and evolve enterprise configurations—compressing months of human work into hours, without sacrificing quality or compliance.

Our founding team built Moveworks (sold for $3B to ServiceNow). We’re in production with Fortune 500 enterprises and seeing explosive demand.

You’ll build the foundation of autonomous enterprise AI — the reasoning, learning, and safety core that makes digital labor real.

🧠 The Role

Part scientist, part systems engineer. You’ll research and build long-horizon AI agents that perform end-to-end enterprise change delivery - from planning through testing and deployment - safely and autonomously.

The challenge: Foundation models can describe changes; they can’t yet own them. You’ll invent new reinforcement learning, planning, and evaluation methods that enable AI agents to handle real enterprise complexity - multi-step reasoning, compliance guardrails, and measurable outcomes.

⚙️ What You’ll Do

Build RL-driven agent architectures:

• Architect agents that plan, execute, and validate enterprise platform changes (ServiceNow, SAP, NetSuite).
• Design hierarchical planners that decompose ambiguous business requests into executable workflows.
• Develop multi-agent coordination between planner, executor, and verifier roles with shared scratchpads and safety guardrails.

Operationalize research into production:

• Implement post-training and fine-tuning pipelines across heterogeneous enterprise data (schemas, ACLs, workflows, integration patterns).
• Build RL and preference optimization loops (offline RL, DPO, PPO, RLAIF) targeting business outcomes, not token accuracy.
• Create simulation sandboxes and gym environments that safely evaluate agents pre-deployment.

Design observability and evaluation:

• Author eval suites that replicate real customer requests, regression tests, and compliance scenarios.