Talk topics

Topics I can speak or run a workshop on, in English or Japanese. Each draws on documented work on Shelly, with the sources linked from the slides.

This page lists topics, not a speaking history.

1. 01

   AI-directed product development

   Who decides, who implements, who verifies. How a product gets built when AI coding agents write the implementation and a person keeps the product calls — and why a single deferred-work ledger matters more than anyone’s memory.

   • Splitting decide / implement / verify
   • Agents reviewing each other, a person choosing which findings matter
   • Cutting a shipped feature two days after launch

2. 02

   Building on-device agents on Android

   Running a terminal, bundled tools and scheduled agents inside one ordinary Android app under Samsung Knox’s SELinux policy, with no root and no proot.

   • Why shebang scripts and am start failed, and what replaced them
   • Shipping tools as native binaries in the APK
   • Agents that fire on a schedule with the screen off

3. 03

   Designing human approval into AI execution

   If agents run unattended, the boundary can’t rest on trust. Every shell command a CLI worker proposes goes through one driver and one gate (whether MCP and app tool calls do is still an open question), approvals come back signed from the phone, and permission levels have explicit limits.

   • Attended versus unattended runs
   • Approvals signed with a key in Android Keystore
   • Finding a gate bypass by testing the gate — and the questions still open

4. 04

   What phone-first development exposed about AI coding

   Verifying on one real device (N=1) changes what “done” means. What a separate “implemented, not verified on device” state reveals, and what it doesn’t.

   • Why passing tests is not the same as working on hardware
   • Freezing a route that worked, over vendor terms
   • What still ran on a PC, and why that matters to say