DevFest Montréal 2026

← Back to the programme
3-5 · Just Can’t Get EnoughEnglish

Beyond Push-to-Talk: Real-Time Fact-Checking Agents with Gemini Live API and Google Search Grounding

Category
AI & ML
Format
Session — 40 minutes
Audience level
Intermediate

Most voice agents wait for you to finish, then answer. That works for a chatbot. It fails on an incident call when someone says the outage started after the deploy and the logs say otherwise, or on a sales call when someone promises a feature that doesn't exist. In this session we build an agent that listens to a live conversation, checks claims while people are still talking, and speaks up only when something is wrong. It runs on the Gemini Live API. We'll cover streaming audio into a live session and turning speech into claims you can check, Google Search grounding for public facts, function calling into your own data like logs or release notes, deciding when an agent should interrupt and when it should stay quiet, and what to do when the session drops or the Wi-Fi dies. You'll leave with a pattern you can reuse for incident bridges, support and sales calls, meetings and live teaching, interviews. Live demo: the agent is listening to this talk, and it's allowed to interrupt me.

Speakers

Aisha Yaseen

AREHSOFT

Aisha Yaseen is a self-taught software engineer. She wrote her first production code at 15, was leading a team of 14 engineers by 19, and has been shipping software for seven years. Today she runs AREHSOFT, an AI consulting firm that builds custom software and AI automation for businesses, and she runs the company itself on AI agents she built, so she knows exactly where they break. On Prompts2Products she has written more than 80 build guides, many on Google AI Studio, Gemini, Firebase and Antigravity, published the same week the features land. She hosts the Prompts2Products LinkedIn Live series, founded Train the Trainers to teach women IT, web development and AI so they can work from home, and organizes Build Night MTL, a Montreal build night that runs on Google AI Studio.