Explore a personal engineering lab for local and cloud AI, agents, persistent knowledge and the systems needed to operate them responsibly.
The Lab combines local and cloud AI, useful automation and inspectable knowledge. This website separates owner-confirmed external capabilities, repository implementation and timestamped public measurements.
Implemented queue, worker and scheduling controls coexist with owner-reported local MLX and knowledge workflows. Current health requires the public telemetry feed; missing measurements are labelled unavailable.
Enforced routing and a larger local inference node are plans. Browser-only preferences are not an operational routing service.
The experiment registry lists reviewed research questions. No dated measurements or experiment conclusions are published yet.
Private local AI, persistent knowledge, reliable multi-agent workflows and human oversight guide the roadmap. Physical AI and robotics are longer-term research, not deployed capabilities.
LIVE — Owner-confirmed, In the code
The control plane. It takes requests, queues work and tracks state.
An API with a PostgreSQL-backed job queue. Workers pull jobs from it rather than being pushed to, schedules live in the same database, and every state change is recorded. PostgreSQL and the API are the only scheduler authority.
The API coordinates work. It does not run models itself, and no code in this repository makes a queued job call Laya or MLX automatically.
LIVE — In the code
The agent CLI used by the fixed research_pack worker workflow.
The Mac worker invokes the supported OpenClaw agent CLI to execute a queued research pack. The API and PostgreSQL retain scheduling and job-state authority.
This execution dependency is not the website's API or scheduler. Its presence does not prove automatic Laya, MLX or vault access.
LIVE — Owner-confirmed
The orchestration layer that decides how a piece of work gets done and which model path it takes.
Owner-described: Laya sits between a request and the model runtimes and, paired with MLX, handles the local path on Apple Silicon.
Laya runs outside this repository. How it connects to the job queue is not verified from the code, no telemetry from it reaches this site, and I do not claim its routing criteria.
LIVE — Owner-confirmed
Language models running on the Mac itself, using Apple's MLX framework.
MLX runs models directly on Apple Silicon and its unified memory. On the current 8 GB machine that means small or heavily compressed models.
Throughput and quality numbers are not published because I have not measured them in a way I would stand behind.
LIVE — Owner-confirmed
Plain-text notes I can read, edit and search on every device. Owner-reported as the Lab's long-term memory.
A folder of Markdown files, owner-described. It is the intended shared memory for me and for agents.
Contents are private. Public pages never show notes, titles or counts that could identify them.
UNKNOWN — Plan only
Models that sense and act in the physical world. A research question, not a project.
Nothing is built. I am reading and thinking about what it would take to do this safely.
No hardware, no prototypes and no timeline.
The current M1 MacBook Pro has 8 GB memory. The M1 Ultra Mac Studio, with target 64 or 128 GB memory, is planned, not live.