Observe
Read host memory, recent trends, local activity, credible LLM growth, and fresh Parallels state.
Product case study · Native macOS
MemoryPilot
A safety-first memory orchestration utility for people running local AI and Parallels virtual machines on the same Mac.
THE PRODUCT IDEA
Local AI should be able to use a powerful Mac without turning every running virtual machine into a manual memory-management chore.
On Apple silicon, local models and virtual machines draw from the same unified memory pool. A heavy model load can rapidly reduce headroom, increase compression and swap activity, and degrade the entire system.
The obvious response—suspend some VMs—is also the dangerous one. A useful product has to know which resources are deliberately managed, whether they are active, what the host is actually doing, and whether the app owns the state it may later reverse.
Design question
How do you make automation genuinely helpful without hiding risk or guessing what the user intended?
02 · Workflow
A control loop that explains itself.
The UI and engine share the same decision language, so monitoring, recommendations, actions, and audit history tell one coherent story.
Reconcile
Match discovered resources to explicit tracking choices and unambiguous app-owned suspension records.
Decide
Apply sustained thresholds, hysteresis, cooldowns, priorities, locks, and protection rules.
Explain
Show the next likely action—or the reason no action is appropriate—before anything changes.
Revalidate
Perform fresh identity, state, and safety checks immediately before a supported command.
Audit
Persist the result locally, record ownership, and use conservative startup reconciliation.
More than a percentage
Physical, used, available, compressed, swap, growth, trend, paging, compression churn, and inferred pressure signals from Mach and sysctl.
Explicitly managed VMs
Per-VM tracking, permanent protection, manual in-use locks, priority, minimum idle time, automatic-resume preference, and temporary exclusion.
Credible process matching
Configurable detection for Ollama, LM Studio, llama.cpp, MLX, and Python transformer workloads, with stricter matching for generic Python processes.
Ownership-aware resume
Only confirmed app-owned suspensions are eligible for automatic resume. Recovery is gradual and defaults to reverse suspension order.
Local, inspectable state
Atomic JSON persistence, bounded audit and metrics history, notifications, diagnostics, policy profiles, menu-bar access, and simulation scenarios.
Disconnected verification
Unit tests and deterministic simulation run without Parallels. Simulation has no route to the real provider.
04 · Safety engineering
What the app refuses to do is part of the product.
No generic power API
The provider has no capability for shutdown, stop, reset, restart, process termination, or arbitrary commands.
Paused + observe-only
Every process launch starts with automation paused; onboarding keeps observe-only enabled.
Exact UUID matching
Fresh JSON discovery must produce one exact, valid match before any supported action can run.
Fail closed
Unavailable CLI, malformed data, duplicate identities, or incomplete reconciliation block automation.
Local by design
No analytics, networking, cloud account, root helper, or guest credentials. Logs hash VM identifiers and omit sensitive arguments.
05 · Architecture
Portable policy, native integration.
Core decision logic stays independent of the UI and Parallels, while the macOS target supplies monitoring, persistence, providers, and focused SwiftUI presentation.
Inputs
Native host metricsMach · sysctl · activity signals
Parallels discoverySupported JSON CLI output
Workload detectionConfigurable local process rules
OrchestratorCore
State + decision enginePolicies · mocks · simulation · explanations
Safety contractsTyped resources · bounded actions
Product
Coordination storePersistence · reconciliation · audit
SwiftUI appWindow · menu bar · settings · diagnostics
Ownership is state, not a guess.
A VM being suspended does not mean the app suspended it. Persisted ownership plus conservative reconciliation prevents automatic resume from adopting a user’s manual decision.
Memory pressure needs context.
Allocation alone is not a crisis. Headroom, swap growth, paging, compression churn, sustained duration, and trajectory produce a more credible signal.
Urgency still needs bounds.
Severe pressure may justify a rapid sequential response, but every candidate is rechecked and every burst has a hard limit.
Incomplete visibility should increase caution.
Without precise window-to-VM attribution, foreground Parallels use protects all running VMs. The fallback is deliberately conservative.
More work
See how MemoryPilot fits the wider systems practice.
Explore infrastructure, local AI, cybersecurity, and troubleshooting projects across the portfolio.
Return to selected work