Build & deploy AI-powered apps — Paige Bailey, Google DeepMind

Paige Bailey, Developer Relations lead at Google DeepMind, delivers a fast-paced, demo-heavy session at AI Engineer Europe 2026 covering the breadth of Google DeepMind's current AI model stack and how developers can immediately prototype with it — almost entirely for free. Bailey opens by orienting attendees around the recent release cadence: over the prior month and a half, DeepMind shipped Gemini 3.1 Flash Live, Gemini 3.1 Pro, Gemini 3.1 Flash-Lite, NanoBanana 2 (image generation and editing), a multimodal embeddings model supporting video, images, audio, text, and code in a single embedding space, Lyria 3 (music generation), Genie 3 (world model building), Gemma 4 (open model family), and Veo 3.1 Lite (video generation). She notes that Augment Code and Replit have already defaulted their entire agent systems to Gemini 3.1 Pro for performance-to-cost reasons.

Bailey works almost exclusively within AI Studio (accessible at ai.dev or aistudio.google.com, free with a personal Gmail account), demonstrating its range of capabilities live. She shows how to link YouTube URLs directly as model input — a 5-minute video costs roughly 27,600 tokens at one frame per second — and demonstrates real-time video analysis using Gemini 3.1 Flash-Lite, which correctly identified dinosaur species including flagging that Pteranodons are pterosaurs, not dinosaurs, using grounding with Google Search for citations. She introduces Compare Mode, running Gemini 3.1 Flash-Lite and Gemini 3 Flash side-by-side on an image analysis task (drawing bounding boxes around green LEGO bricks using Python via code execution), with the entire task completed at well under a fraction of a penny.

Code Execution — a sandboxed Python environment with pre-installed data science libraries invocable as a tool call — is highlighted as an underutilized feature. The session demonstrates using it to attempt SVG generation from a LEGO brick image, with Gemini 3.1 Pro (high thinking mode) reasoning through perspective, grid calculation, and color definitions before producing a rough but recognizable output. URL Context is shown as a lightweight retrieval alternative: the model can ingest public URLs, answer questions grounded in their content, and cite sources inline without requiring a vector database.

Gemini Live — real-time screen sharing, video feed, and audio conversation with the model — is demonstrated including multilingual output (Italian and Spanish verified live by audience members), Texan-accent poetry generation, and finger-counting via webcam. Bailey describes practical applications: bank entryway kiosks responding in a customer's native language, AR-assisted navigation, and robotics orchestration via the open-source Pupper robot (3D-printed, Raspberry Pi-based) using Gemini Live for object detection and action planning.

AI Studio Build is shown creating a full-stack bookshelf cataloger app from a voice-dictated prompt — the app uses Gemini 3.1 Pro, Firebase Firestore for persistence, and Google OAuth — deployed and shared live within minutes. Genie 3, DeepMind's world-model system (composed of NanoBanana, Veo, and Gemini), generates a navigable Big Bend National Park scene with LEGO rock formations and a pink ostrich character with rocket blasters; Bailey clarifies it generates pixels dynamically per frame rather than 3D meshes. Veo 3.1 Lite produces Warriors-themed vegan food truck stock footage at 720p from a Gemini-expanded prompt. Lyria 3 generates a Spanish-language electronic track in real time. Bailey closes with a note that Gemma 4 models are available via AI Studio APIs for free testing, run on Pixel devices, and are integrated into Chrome, with Apache 2 licensing for self-hosted fine-tuning.