Taught CI/CD to students with one deleted comma.
Deploy by hand: the site broke, with no record of who shipped it. Deploy through the pipeline: tests failed, deploy skipped, production untouched.
Same bug. Different kitchen. 🍛
When both answered on a real run:
rules said "possible curiosity 0.50",
Claude said "possible play invitation 0.66".
Disagreement is fine.
Showing both is the honest part.
Hardware + LLM projects have a problem:
the internet drops, the API is slow,
and your device just sits there.
PawBridge used two tiers
so the device always answers. 🧵
Tier 1: local rules on the laptop.
Ultrasonic trigger (within 30 cm)
+ accelerometer (taps, pawing, stillness)
→ an answer instantly, no internet needed.
Ran an AI agent with Wi-Fi OFF at a talk in Madurai.
Gemma 4 on Ollama, on my laptop: sorted files, filled a form, did time math's from the agenda.
Then "plan a 3-day trip to Madurai": too open-ended, so it handed off to Claude.
LLM 🐘 🆚 SLM 🐈
Code: https://t.co/nnU1dLby3p
First multi-agent system in Copilot Studio. What surprised me: Al helped with the planning, but the Power Automate flows were still manual work.Error handling took more thought than the agents.And the compliance docs raised more questions than the build did.
What hurt you most?
Our dog-behaviour device had one rule enforced in code: every answer must start with “possible”.
“Possible play invitation, 0.66.”
Never “your dog wants to play.”
If an AI can’t be sure, make it impossible for it to sound sure.