ASTRA-IDE is built on seven contributions. Each one is evaluated on a real public dataset, and every headline number below comes from a trained artifact committed in the repository. Each card links to the paper it implements and the official dataset, so any claim can be traced to its source.
A deep reinforcement learning agent places each workspace across heterogeneous nodes.
PPO policy trained on the Google trace; outperforms random workspace placement
Per-workspace syscall and resource capture in the kernel with near-zero overhead.
HashPipe heavy-hitter detection, O(d*m) memory regardless of process count
An LSTM forecasts session demand and adapts container keep-alive to cut cold starts.
cold-start reduction vs a fixed 10-minute keep-alive window (N-RMSE 0.17)
A risk model picks the cheapest safe isolation tier; a graph model flags exploits.
on LID-DS CVE traces, above STIDE (0.75) and a frequency baseline (0.73)
Karmada spreads workspaces across clusters with automatic failover.
workspace pods reschedule when a member cluster is taken down
Placement and deferrable work follow live grid carbon intensity.
CO2 reduction at a 24-step deferral budget on real UK grid data
Multiple users edit the same file with conflict-free CRDT synchronization.
order-independent merge verified on a real keystroke trace
Evaluation harnesses live under benchmarks/; trained artifacts and their metrics live under ml/*/artifacts/ in the repository.