The consultant will build an AI Agent that turns a diagnosis into a proposed code change, opened as a pull request for a human engineer to review (the agent proposes, never merges) - related to a reliability capability product that detects faults in the data platform and triages them into a diagnosis.What you will do

• Define the narrow class of faults the agent may propose fixes for, and enforce that boundary. The allowlist is a safety control, not a backlog.
• Retrieve the right repository context for a given diagnosis.
• Generate the proposed change and verify it locally before anything is opened.
• Open a pull request carrying full provenance: what was detected, what evidence was used, what changed, and why.
• Design the reviewer experience, including rejection. Rejection feedback is the primary learning signal and most implementations ignore it.
• Enforce technically that the agent cannot merge, under any configuration.
• Measure acceptance rate and correctness, and report the numbers that do not flatter the work.

What we need from you

• You have built LLM agents that act on real systems and evaluated them, not just demonstrated them.
• Python.
• Git and GitHub in depth: GitHub Apps, fine-grained permissions, Actions, branch protection.
• The ability to treat the safety boundary as a design problem rather than a policy statement.
• Comfort producing work that is reviewed by engineers who did not ask for it.