← Observatory
ahmedmshazly /

adaptive-pipeline-orchestration

FLAGSHIP
testedreproducible

RL vs. rule-based data-pipeline orchestration under one shared utility — a rigorously hardened negative result.

A research study that reports what actually happened — the clever method didn't win — and lets anyone re-run every number.

  • Python 85%
  • TeX 13%
  • Makefile 2%
languagePythoncreatedMar 2026updatedJun 2026size8.3 MBlicenseMITView on GitHub ↗

Why this exists

Graduate research asking whether a self-learning agent can orchestrate data pipelines better than fixed rules. The honest headline is that on the original environment it can't — and the repo documents the skeptical-reviewer pass that made that claim defensible: disjoint seed pools, a reward-identity fix pinned by regression tests, and environment variants that show where learning does start to win.

Start reading here

what I'd do differently now

I'd stress-test the environment before the agent: the benign default made a trivial policy optimal, and most of the hardening pass was earning back the right to make any claim at all.

Open to internships & contract work · full-time from 2027

Get in touch →