2026
Ask Judy
A multi-agent second opinion on whether an online friendship is safe.
- Role
- Concept, agent design and build
- Team
- Five, from UCL
- Timeline
- Oct 2026, one-day hackathon
- Tools
- Figma, Claude, Supabase
Deployed prototypeLive prototype
Problem
People who are isolated online find it hard to tell a friend from a scam.
Approach
Three agents read a shared chat, judge it against the user's own boundaries, and check each other.
Outcome
A deployed, working prototype built at Accel's AI Innovate hackathon in London.
I shaped the concept and built the agent layer.
State your role first: concept, PRD, agent plan, flows, agent code. Name what teammates owned (UI, database).
Two interviews changed what Judy should never do.
What the two interviewees said and which design decisions each finding led to.
Four answers instead of two, including 'I'm not sure'.
Two states
Why it was too confident.
Four statesChosen
Why admitting uncertainty builds trust.
A reader, an analyser and an independent checker.
Shipped on the day, and what the judges said.
Deployed and working; didn’t make the top eight. Summarise the judges’ feedback honestly.
