Aster — Distributed AI Research Scientist
Built a hackathon-stage research orchestration MVP that turns an ML question into a literature brief, falsifiable hypothesis, bounded experiment plan, worker execution, analysis, and an evidence-labelled Markdown report.
Related Projects
SYSTEM ARCHITECTURE
Offline-demo mode uses deterministic agents and simulated workers where external services are unavailable.
Problem
ML experimentation spans literature review, hypothesis design, experiment planning, compute coordination, and result interpretation; Aster connects these stages in one inspectable workflow.
Architecture
Research question → literature and hypothesis agents → human-reviewed experiment plan → simulated workers or optional DMLF bridge → result analysis, run timeline, and reproducibility manifest.
Challenges
- Kept research plans bounded and added a critic step with explicit caveats and human-approval requirements
- Integrated an asynchronous worker protocol with job submission, status polling, progress updates, and metric collection
- Connected the orchestration layer to the separate Distributed ML Training Framework
- Recorded run provenance and distinguished simulated results from real execution
Benchmarks
The default offline workflow uses deterministic agents and simulated workers; its preset metrics are not benchmark evidence. The repository documents a two-node DMLF synthetic execution smoke test and an MNIST path, but benchmark claims should be made only from a completed run with documented data provenance.
Lessons Learned
- Research orchestration needs explicit review gates and inspectable intermediate artifacts
- Distributed execution and scientific evidence are separate concerns
- Run manifests and clear simulated-versus-real labels make results easier to audit