01 Research Bench
A public notebook for work in motion. Hypotheses change. Failed tests and discarded directions stay visible.
Active
One self-hostable gateway can reduce credential exposure and tool noise without hiding useful capabilities.
Juraguard consolidates remote MCP servers behind one endpoint, keeps credentials server-side and exposes compact search and call tools.
Active
Durable personal context matters more than sending the agent every available document.
Personal Agent is testing selected durable context alongside on-demand MCP tools rather than one giant prompt.
Active
Voice lowers friction, but every non-read-only action needs an explicit native approval step.
The iOS experiment combines a voice-first client, connected tools and approval before actions that change external systems.
Active
Home Assistant over MCP can provide one useful control surface if permissions remain narrow and visible.
Building the MCP integration against my own Home Assistant setup before making broader claims.
Discarded for now
A narrow model may learn repeated personal routines without needing a general-purpose agent.
No useful proof yet. Parked rather than presenting speculation as a project.
Active
Use AI to research costly operational problems, while people own targeting, judgment and every message.
Pyango is testing a problem-first workflow. Private account data, numbers and outreach details stay private.
Failed tests
A signal deserves paper trading only after walk-forward validation, realistic costs and stability checks.
Candidates that looked good in a backtest failed later gates. Rejection is the result, not a hidden inconvenience.
Active
Memory should be deliberate, reversible and separate from authority to act.
Comparing durable context in Personal Agent with explicit approval and server-side tool boundaries in Juraguard.