← Applied AI engineering

Public personal engineering

SEG

Agent evaluation and orchestration

Builder and maintainer
2026

  • Python
  • DAG scheduling
  • Evaluation
  • Staged repair
  • CI

The problem

An evaluation result is not useful if a failed or skipped check can silently become success. Repairs also need to be checked before they change the target files.

What I did

  • Track dependency, failure, timeout and skipped states explicitly.
  • Run versioned control and treatment scenarios; record harness failures as invalid trials.
  • Stage repairs, re-evaluate them and check changed inputs, paths and recovery before edits.

The result

Built and maintain the evaluation runtime, with runtime and repair tests in Windows and Ubuntu CI.

Scope and boundaries

Personal engineering, not a client production deployment. Repository evidence and test results should be read at the relevant revision.

Explore the public repository

Portfolio summary updated September 2026.

Talk to me about similar work