Causal self-healing
Fix hit rate
higherInjected config and package regressions across 64 service traces.
ML shouldn't be exclusive to the application layer.
DracOS is an operating system concept that learns from workloads, failures, and runtime behavior to make the machine more adaptive without asking users to become system tuners.
* pre-seed
Results.
Mean +/- standard deviation across lab runs. Ubuntu 24.04 is the fixed-policy baseline.
Causal self-healing
Recovery time
lowerCounterfactual rollback ranking against scripted triage.
Kernel tuning
4k randread
higherfio 4k random read on ext4 NVMe, warm page-cache window.
Kernel tuning
SQLite rebuild
lower1.6M-row import with adaptive writeback thresholds.
Roadmap.
DracOS was founded by two students from ETH Zurich exploring how ML can move from application features into the operating system substrate itself. The initial prototype focuses on learning kernel parameter and cache policy choices, plus causal self-healing for reversible fixes. In the future we plan to tackle uncertainty and integrated AI memory.
Learned kernel tuning
DracOS studies workload shape and runtime signals to recommend kernel parameters, cache policies, scheduler settings, and power profiles without manual sysctl tuning.
Causal self-healing
A causal graph connects processes, configurations, updates, and failures, then ranks the reversible change most likely to fix the system.
Uncertainty fallback
Learned policies should expose confidence and fall back to deterministic Linux behavior when the model is uncertain.
AI memory
A permissioned memory service shared between local models and applications, with provenance, expiration, and privacy boundaries.
Request a prototype or set up a call.
DracOS is early stage. Reach out if you want to see the prototype, discuss kernel-level ML, or follow the work as it develops.