Oct 2026 · Agent tools and evaluations, team project
PAilot
2nd place (Public) and Best Agents Use Award, NUS-ISS Show Me Your Agents Hackathon
An agent that reschedules field-service technicians after disruptions, with adversarial evals that prove unsafe runs change nothing.
PAilot is an AI-assisted dispatch desk. When a technician is delayed or a job changes, it helps a coordinator rebuild the schedule.
What I built
- Typed LangGraph tools that validate inputs and workflow state before anything runs. Consequential changes need a human to approve them.
- Normal and adversarial evaluations covering prompt injection, invalid tool calls, and stale planning data. Each eval checks which tools ran and verifies that unsafe runs left the schedule unchanged.
Why it matters
An agent that edits real schedules has to fail safely. The evals were how we showed the judges it does.