Senior Machine Learning Engineer, Agent Eval Platform

The Role

Moveworks' AI agents don't just generate text — they act. They plan, call tools, and change real state in enterprise systems on behalf of 5.5 million employees. Building agents is no longer the bottleneck. Improving them is: knowing whether a change made the agent better, why a conversation failed, and which part of the system should be fixed.

You'll build the loop that makes our agents improve themselves — finding failures in real use, deciding which part of the agent harness is responsible, checking every proposed fix on cases the fixer never saw, and turning each fixed failure into a permanent test. We improve the system around the model: instructions, tool descriptions, routing, knowledge and memory. We don't train the model itself.

The work is blend of ML and systems engineering. You'll decide whether a signal is trustworthy enough to act on, and build the infrastructure that produces that signal reproducibly across hundreds of customer organizations.

This isn't a model-training role, and it isn't a testing role. It's applied ML at a point where the methodology genuinely isn't settled: LLMs judging LLMs is an open research problem, and we're working it against agents that take real, irreversible actions in stateful, multi-tenant enterprise environments.

What you get to do in this role:

Judge design and calibration

• A shared base judge with per-item rubrics expressed as configuration next to the dataset — so eval authors express intent, rather than forking a prompt per eval
• Splitting the problem correctly: deterministic validators for checkable world state ("was the ticket created, with the right item, routed to the right approver?"), and an LLM judge for the parts that are genuinely fuzzy — was the clarifying question appropriate, was policy followed, was the path efficient
• Scoring that reports its own confidence, so uncertain judgements route to a human instead of quietly becoming trusted signal
• A standing calibration loop against human-labeled trajectories, run in partnership with our annotation team — they own the human labeling, you own the calibrated judge artifact. How consistently humans agree with each other sets the ceiling on how good any judge can be, so raising that ceiling is part of the job
• Guarding against correlated blind spots: our user simulator and our judge are both LLMs, and they can be wrong in the same direction
• Offline↔online divergence: when simulation and production disagree, being the person who can say why, and keeping the suite re-seeded from new production failures so it can't quietly overfit
Self-improving agent harness

This is where the pillar is headed, and a large part of why the seat exists. You'll build the components that let the agent harness get better from its own use, without changing the model underneath.

• Finding what to improve: surfacing where the agent falls short in real use, including failures that look like success
• Synthetic data and environments: generating scenarios and the seeded enterprise worlds they run in — with expected outcomes and checks — from real failures or ahead of what customers ask for, and making sure what gets generated is realistic, hard enough, and doesn't leak the answer
• Deciding where a fix belongs: understanding why the agent failed and which part of the harness should change: its instructions, tools, routing, knowledge or memory
• Proposing improvements: optimizing those components against simulation rather than against production traffic
• Checking before anything ships: verifying every proposed change on cases it wasn't tuned on, catching regressions an average would hide, and keeping the part that proposes a change separate from the part that judges it
• Evals that grow: every fixed failure becomes a permanent test, so the suite gets stronger the more the agent is used
• Safe across many customers: learnings stay scoped to one organization where they must, and admins approve anything org-specific
• The infrastructure that runs it: a loop that runs reproducibly, at scale, across hundreds of customer organizations
• The guardrail that keeps it honest: end-state outcomes are what we hold the agent to

Function: Engineering
Experience Level: Not Applicable