A scheduler should learn from the decisions it makes.
Choosing the lowest predicted latency or cost is only a starting point. Predictions can be uncertain, workload mixes can shift, and the outcome of the unchosen placement is usually missing. Adaptive AI Orchestrator investigates whether selective counterfactual exploration can reduce scheduling regret while respecting service-level objectives.
From choosing a strategy to placing a job.
Enterprise AI Workload Intelligence studies which model, tool, or execution strategy should handle a workload. This project carries the allocation question into compute scheduling: where should an inference job run given latency, cost, capacity, reliability, and uncertainty?
It is a related experiment with a different decision unit. Its emphasis is placement feedback, shadow alternatives, drift detection, and a decision history that can be inspected and replayed.
Place, observe, compare, and recalibrate.
- Represent jobs, compute nodes, predictions, and outcomes with typed models.
- Choose risk-aware placements using uncertainty bounds rather than point estimates alone.
- Select a shadow candidate when the potential information justifies its cost.
- Compare the chosen outcome with the observed alternative and watch latency and quality drift.
- Preserve decision lineage and inspect assignments in an interactive decision table.
Compare policies on identical seeded workloads.
The deterministic harness compares FIFO, least-loaded, predicted-best, and an uncertainty-aware counterfactual policy on the same synthetic workloads. It records scheduling regret relative to the best feasible observed outcome, success, service-level violations, and realized cost.
Benchmark runs accept a seed and job count and produce table, JSON, or Markdown reports. This supports reproducible policy comparisons without treating simulated performance as measured production gains.
A research scheduler with a measurable next step.
The current milestone includes placement logic, shadow-candidate selection, drift detection, the inspection interface, and an evaluation harness in a Next.js application. GPU utilization and recovery after workload drift are among the planned evaluation dimensions.
The repository does not establish production SLO compliance or a measured improvement on a live compute fleet. Those claims require observed workloads, controlled rollout, and validation against real resource behavior.