# Build & deploy AI-powered apps

Paige Bailey, Google DeepMind | AI Engineer Europe 2026 | 1:03:20

Source: https://www.youtube.com/watch?v=G_bHFmEAarM
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/build-deploy-ai-powered-apps
Published: 2026-04-29
Tags: code-generation, deployment, multimodal, tool-use

## TL;DR
- AI Studio lets developers test Gemini models with video, images, audio, URLs, search grounding, code execution, and generated code from the same interface.
- AI Studio Build can create and deploy applications with Firebase databases, Google authentication, persistent user data, and public sharing from a natural-language description.
- Google's model family covers different needs, from Gemini 3.1 Pro for complex app generation to Gemini 3 Flash and Flash-Lite for faster, lower-cost work, plus Gemma 4 for downloadable open-model use.

## Summary
Paige Bailey gives a demo-heavy tour of Google's AI development tools and models. She starts with AI Studio, showing video analysis from a YouTube URL, search grounding, URL context, model comparison, and sandboxed Python code execution. Small models can handle tasks such as drawing bounding boxes around objects at very low cost. Gemini Live supports screen, video, and audio input, natural conversation, custom functions, search grounding, language changes, and style changes. AI Studio Build turns a plain-language description into a deployed app. Bailey builds a bookshelf cataloger with Google login, Firebase storage, search grounding, and persistent results. She also demonstrates Genie 3 world generation, Nano Banana 2 image editing, Veo 3.1 Lite video generation, Lyria 3 music generation, and Gemma 4 multimodal models. She is clear that Genie 3 generates pixels rather than reusable 3D meshes, and that some features still require additional work or are not yet available through an API.

## Key ideas
### Google's model family covers different costs, sizes, and media types
[01:18](https://www.youtube.com/watch?v=G_bHFmEAarM&t=78s)
Bailey describes a fast-moving set of models from Google DeepMind. Gemini 3.1 Pro is the largest and most expensive option, Gemini 3 Flash is the common production workhorse, and Gemini 3.1 Flash-Lite targets speed and low cost. Nano Banana 2 handles image generation and editing, Lyria 3 generates music, Genie 3 creates interactive worlds, Veo 3.1 Lite generates video, and Gemma 4 belongs to the open model family. She also describes an embeddings model that puts video, images, audio, text, and code into one embedding space, allowing a search for cat-related content to return cat videos, images, sounds, and books.

### Gemini can take and produce more than text and code
[03:35](https://www.youtube.com/watch?v=G_bHFmEAarM&t=215s)
Gemini accepts video, images, audio, text, and code as inputs, and can produce text, code, audio, images, and other multimodal outputs. Through the APIs, developers can pass PDFs with embedded images, different kinds of video, and different kinds of audio. Bailey presents this as a practical difference from models that mainly accept text and code and return text or code. The flexible input formats make it possible to build applications around media rather than treating media as an isolated preprocessing step.

### AI Studio adds current information through tools and context
[05:33](https://www.youtube.com/watch?v=G_bHFmEAarM&t=333s)
AI Studio lets users turn on structured outputs, code execution, function calling, Google Search grounding, Google Maps grounding, and URL context. Bailey explains that a model cannot answer about events after its training data cutoff without access to search or retrieval. In her video example, Gemini analyzes five minutes of a YouTube dinosaur video, creates a timestamped table of dinosaur types, and uses Google Search to add facts and citations. AI Studio also provides the code needed to reproduce the experiment in Python, TypeScript, or Java.

### Small models become more useful when they can call tools
[13:06](https://www.youtube.com/watch?v=G_bHFmEAarM&t=786s)
Bailey compares Gemini 3.1 Flash-Lite with Gemini 3 Flash on an image task. Both models use code execution to write Python that detects green LEGO bricks and draws bounding boxes. Flash-Lite gets the answer quickly, while Gemini 3 Flash reaches the same result more slowly and at a slightly higher cost. Bailey says both remain well under a penny for the task. She recommends trying smaller models with tools such as code execution, since the tools can help them solve image analysis and data tasks without requiring a larger model.

### Gemini Live supports real-time interaction with screens, cameras, and audio
[18:29](https://www.youtube.com/watch?v=G_bHFmEAarM&t=1109s)
Gemini Live can receive a shared screen, video feed, or audio and respond dynamically. Bailey uses it to describe a browser page full of LEGO images, repeat the answer in Italian, report what is happening in London, write a poem in a Texan accent, and identify the number of fingers shown to the camera. She says developers can also add custom function calls, automatic function responses, and Google Search grounding. System instructions can set a required language, dialect, accent, style, or tone. The same pattern could support a bank display that answers visitors in the language they prefer.

### AI Studio Build can create a working app with authentication and data storage
[25:01](https://www.youtube.com/watch?v=G_bHFmEAarM&t=1501s)
Bailey demonstrates AI Studio Build by describing a bookshelf cataloger through speech input. The app accepts a bookshelf photo, identifies book titles and authors, uses Google Search grounding for descriptions and categories, asks users to sign in with Google, and stores the results in a database. Gemini 3.1 Pro plans and writes the application, including TypeScript, CSS, Firebase configuration, and database rules. The finished app accepts a photo, adds the detected books, persists them after sign-out and sign-in, and can be made public with a shareable link. Bailey says deployment to Cloud Run is available through a one-button action.

### Genie 3 generates navigable pixels rather than reusable 3D game assets
[29:39](https://www.youtube.com/watch?v=G_bHFmEAarM&t=1779s)
Project Genie uses descriptions of scenes and characters to generate worlds that users can explore with keyboard controls. Bailey creates Big Bend National Park in Texas with LEGO rock formations, a quadruple rainbow, and a pink ostrich with a rocket blaster. The system produces changing views and responds to actions such as moving through the scene or hitting objects. She explains that Genie 3 has no Unity or Unreal Engine physics system behind it and generates each frame dynamically. It does not create 3D meshes or game assets. Turning its output into reusable 3D assets requires other techniques.

### Open models and generative media can be tested in the same workspace
[49:49](https://www.youtube.com/watch?v=G_bHFmEAarM&t=2989s)
Bailey presents Gemma 4 as a downloadable open model with an Apache 2 license. It supports multimodal understanding, can run on mobile devices in its smallest versions, and is available to try through AI Studio. She also uses Nano Banana 2 with image search grounding to place a dog in a natural park holding a Celsius can. Veo 3.1 Lite generates 720p video in several aspect ratios and durations, while Lyria 3 creates music from a style, subject, and language request. AI Studio exposes the code for these experiments so developers can move from a gallery example to their own application.

## Notable quotes
- "It gives you the option to to kind of create and deploy, and to share, a whole spectrum of apps." (25:01)
- "This super tiny model worked really really fast, wrote the Python code to pull in the image, to analyze it, and to define the bounding boxes." (15:07)
- "It is just the pixels." (37:01)
- "If you need a lighter-weight model accessible via an API that you can work with for free, or if you need a model that you can download, use on your own infrastructure, fine-tune, and run for free with an Apache 2 license, the Gemma 4 models are an incredible option for you to try." (52:15)

## Tools & references mentioned
- Google DeepMind
- Gemini 3.1 Flash-Lite
- Gemini 3.1 Pro
- Gemini 3 Flash
- Gemini Live
- Gemma 4
- Nano Banana 2
- Lyria 3
- Genie 3
- Veo 3.1 Lite
- AI Studio
- AI Studio Build
- Firebase
- Firestore
- Cloud Run
- Google Search
- Google Maps
- Vertex
- OpenClaw
- MediaPipe
- Pupper
- Raspberry Pi
- NumPy
- SciPy
- scikit-learn
- Replit
- Augment Code
- v0.dev
- Lovable
- World Labs
- Fei-Fei Li

## Who should watch
- You want to turn a prototype idea into a small application with login, a database, and a shareable deployment without setting up the whole stack by hand.
- You need to choose between Google's larger, faster, cheaper, open, or media-generation models for an application.
- You are building multimodal features around video, screen sharing, image analysis, live audio, generated video, music, or interactive worlds.

## Related talks

- [Building in the Gemini Era](https://aietalks.com/talks/building-in-the-gemini-era) (Kat Kampf & Ammaar Reshi, Google, 17:57)
- [Prompt to Pipeline: Building with Google's Gen Media Stack](https://aietalks.com/talks/prompt-to-pipeline-building-with-googles-gen-media-stack) (Paige Bailey, Guillaume Vernade & Ian Valentine, Google DeepMind, 1:54:35)
- [Building Conversational Agents](https://aietalks.com/talks/building-conversational-agents) (Thor Schaeff & Philipp Schmid, Google DeepMind, 1:47:34)
- [AI Engineering with the Google Gemini 2.5 Model Family](https://aietalks.com/talks/ai-engineering-with-the-google-gemini-2-5-model-family) (Philipp Schmid, Google DeepMind, 1:44:51)
- [How Google DeepMind Is Researching the Next Frontier of AI for Gemini](https://aietalks.com/talks/how-google-deepmind-is-researching-the-next-frontier-of-ai-for-gemini) (Raia Hadsell, Google DeepMind, 20:37)