Always on
The server owns the assistant
Triggers, memory, permissions, integrations and logs remain available even when every laptop is closed.
Always-on AI · Homeservers · Useful automations
Building a personal AI that is always on.
The home server does not need to be the smartest model in the room. It needs to stay awake, own the context and permissions, run dependable workflows, and choose the right local or cloud model for each job.
The architecture
Keep the control plane at home.
The assistant's durable state lives on the server. Inference is a replaceable component, not the whole product. That creates a useful system without pretending a storage-focused NAS is a frontier-model datacenter.
Hybrid inference
Use the right brain for the job
Rules first, small local models for bounded private work, and approved cloud models when the task needs more capability.
Useful automation
Judge completed jobs
Measure corrections, failures, latency, cost and maintenance—not whether every component carries an AI label.
Why the server matters
Always available. Selectively intelligent.
Context stays local by default. Send only the minimum approved context when a job genuinely needs a cloud model.
Own the boring parts. Backups, logs, updates, and recovery matter more than a flashy demo.
Automate with an escape hatch. Every workflow needs visibility, limits, and a way for a human to take over.
On the workbench
Projects, not promises.
The lab will evolve as the server does. These are the threads I'm actively building or carrying forward.
01
Building
02The always-on home assistant
Using a UGREEN homeserver as the control plane for context, memory, permissions, automations, and deliberate model routing.
Exploring
03An automation control plane
A simple way to see what ran, what failed, what an LLM decided, and where human approval still matters.
Available
Coding-agent instruction evaluator
The original Sharper project: replay repository tasks and test whether instruction changes actually help an agent.
Latest lab notes
Build logs and explainers.
Home assistant architecture
Browse every lab note →An always-on AI home assistant: what should run on your server?
A practical architecture for an always-on personal AI assistant: keep memory, permissions, triggers, and automations at home while routing inference to the right local or cloud model.
Read the note →Build logWhy I’m building a local AI homeserver
The plan for turning a UGREEN server into a useful base for local LLMs, automations, and self-hosted tools—without pretending the hard parts disappear.
Read the note →AutomationDesigning LLM automations that remain useful after the demo
A practical framework for choosing home automations, constraining the model, and making every run observable and recoverable.
Read the note →Compare notes
What should an assistant do at home?
Share the recurring household job, homeserver constraint, or privacy boundary you want turned into a measured experiment.