Grilled
A fully client-side mock interview and exam prep tool built on the Gemini Live API — real-time bidirectional audio/video, adaptive questioning, and AI-generated study materials. Built at the {Tech: Europe} London AI Hackathon for Google DeepMind.
Grilled was built at the {Tech: Europe} London AI Hackathon in April 2026, targeting Google DeepMind's "Real-Time Conversational Agents with Gemini 3" track.
The challenge was to build a conversational agent that felt like a real interviewer — one that could listen, watch, adapt its questions on the fly, and give genuinely useful feedback afterwards.
The core is the Gemini Live API, piping bidirectional audio and video over WebSocket. The agent interrupts on errors, encourages during hesitation, and escalates when performance is strong — all in real time with no backend.
After each session, Gemini Flash generates a three-tier gap report (strong / needs review / priority) with a readiness score, and the image generation model produces visual study aids for weak areas.
Shipped a working demo in under 48 hours. The dual-mode design — Interview Prep for mock interviews and Exam Buddy for oral exams — covered a wider surface than most hackathon entries.
The fully client-side architecture meant zero infrastructure cost and instant deployment.