More beta testers are coming on board this week. The Aide AI-powered analyst is already proving its worth where local governance needs it most. ๐ https://t.co/550lvsKhDS
Aide just hit its first local council meeting โ and crushed it. In head-to-head testing against existing AI models by a governance tester, Aide proved MUCH better at forensic assessment. ๐งต
Shoutout to @dzhng for the assist. Workspace resolved and a reminder that agents make things more complex than they need to be. After rolling back a few over-engineered versions and tuning Vesper (our main agent), we got a smooth sign-up flow running.
That's this weekโs update for Building with @duetchat's Summer of Tokens:
Next up is self-service document ingestion and sign-ups. This beast is nearly finished ๐
Aide just had its first two encounters with the real world, and the results speak for themselves.
759 pages of council agendas. 8 hours of debate. Endless clutter.
Hereโs how a dispassionate AI brief shifted the course of two committee meetings ๐
Knowing where to look, armed with probing, low-outrage questions, puts decision-makers on a completely different footing.
Our beta tester is already sharp and AI-savvy but Aide's dispassionate approach shifted the dynamic from loud opinions to thoughtful, fact-anchored scrutiny.
Biggest takeaway so far? Fix the slowest bottleneck.
AI models are blazing fast, but when the surrounding process forces human intervention to feed them tokens, everything grinds to a halt.
Weโre building AI that holds the line on accuracy.
Local government councillors face 700+ page agenda packs with real ratepayer money on the line.
We pointed Aide โ the AI decision support tool we're building with @duetchat's Summer of Tokens at its first live agenda (226 pages).
Hereโs what we learned ๐
When Aide hit a technical wall, it didn't hallucinate. It flagged: "Held for human review."
That might seem like a tech fail. Actually, it's a win. Compute is cheap โ the real value lies in engineering verification, strict checks, and built-in discipline.
For Aide, AI must be impeccably trustworthy. If it gets a $10M decision wrong, itโs dead. Weโre repeatedly testing the gap between what the system says and what the source supports.
More to come as we keep building with @duetchat's summer of tokens!
Local councillors read 1,000+ page agendas to make $10M decisions. This week, we built an AI Analyst (Project Aide) to help themโand quickly learned that teaching AI principles instead of rules is anything but straightforward. ๐งต๐
Even seeded with 4 months of context, our first iterations failed. The AI repeated instead of analyzing and confused proposals with adopted decisions.
Progress was slow until we used our veteran AI agent, Vesper, to critique the new AI and teach it to self-correct.
Thanks to @duetchat, weโre building Aide: decision support for NZ local councillors.
Aide reads 100s of pages and presents decision impacts, assumptions, and contradictions so councillors can see more clearly.
The foundation is in. Next up: the Analyst agent.
@duetchat, meet Vesper โ AI employee for the PARALLAX crew. He's currently crushing it for strategy, deep analysis, SEO, and positioning work... and when he can have his own avatar, he'll look something like this :)