From idea to production: Building AI apps with GDE Tomek Porozynski
Summary
In this fireside chat recorded live at Google Cloud Next '26, Aja Hammerly (Director of DevX AI at Google) interviews Tomek Porozynski (also referred to as Tomek Wierzchowski in the transcript), a Google Developer Expert for Cloud AI and Staff ML Engineer at Deep Sense AI, about his journey building a multi-voice AI audiobook generator using Google's Gemini APIs.
Tomek describes the core technical challenge of the project: taking plain text and converting it into an audiobook with distinct voices for each character. The workflow involves several steps — first, using Gemini to parse the story and determine whether the narrator is a first-person or third-person voice; second, identifying all characters in the text; third, assigning distinct voice profiles to each character; fourth, splitting the text into chunks by speaker; and finally, synthesizing audio via the Gemini Text-to-Speech API. Tomek notes that special effects (ambient noise, sound design) remain on the to-do list.
A key architectural insight Tomek shares is a hybrid serial/parallel processing model. Some steps must run linearly — character extraction and text segmentation must complete before any audio is generated — but once the text is split, multiple TTS API calls can be made in parallel simultaneously, significantly speeding up production time.
The project is open-sourced and published as a Google Colab notebook (github.com/ontaptom/cairo), intentionally structured so that each cell is independently runnable and annotated, enabling developers to understand, modify, and extend the workflow.
Tomek discusses the critical role of Agent Skills in the Gemini CLI for keeping AI agents informed about the latest model names and API interfaces that postdate the model's training cutoff. He notes: "AI technology evolves so fast that the model isn't always aware of the latest updates. That's where Skills come in — they give the model the guaranteed link to the most up-to-date resources."
On developer workflow, both Tomek and Aja emphasize a "plan first" philosophy — spending significant time in a long conversational brainstorm with the AI agent before writing any code, discussing tech stack options, evaluating the idea, and stress-testing the architecture. Tomek describes an 80/20 rule where the brainstorming phase is the most critical investment.
For troubleshooting, Tomek highlights the Antigravity IDE's ability to spin up a Chrome browser, click through graphical interfaces, and self-debug, enabling near-fully automated error correction loops.
On onboarding new AI developers, Tomek recommends starting small and resisting AI FOMO. For those with coding skills, he recommends Antigravity as an IDE. For non-coders, Google AI Studio's build mode can generate and deploy a working app in minutes, with Firebase Auth for authentication, Firestore for stateful data, and Cloud Run for production hosting with custom domain assignment.
Key Takeaways
- •Invest heavily in the brainstorming phase — have a long conversation with your AI agent about tech stack, architecture, and tradeoffs before writing a single line of code.
- •Use Agent Skills in Gemini CLI to ensure your agent always references the latest API documentation and model names, especially for fast-moving APIs that may have changed after the model's training cutoff.
- •Design AI pipelines with hybrid serial/parallel execution: run steps that depend on each other in sequence, then parallelize independent calls (like batch TTS synthesis) to improve performance.
- •Don't give up after the first failed prompt — ask the agent to self-correct, rerun, and iterate. Tools like Antigravity can automate browser-based debugging almost entirely.
- •Start small: build the core logic in Google AI Studio, then layer in Firebase Auth, Firestore for state, and deploy to Cloud Run when ready for production.
- •Open-source your experiments as Colab notebooks so others can run individual cells, understand the code, and adapt it — this is a highly effective teaching format.
- •Resist AI FOMO — the constant stream of new models and tools is distracting. Pick a tool that fits the task and build with it rather than chasing every new release.
- •For developers building AI apps: Antigravity IDE is a strong choice. For non-developers wanting to prototype: Google AI Studio's build mode generates and deploys apps in minutes.
- •Proof-of-concept first, then scale — validate each component chunk individually before combining them into a larger integrated solution.
Transcript
Hey folks, I'm Audrey Hamel, from the builder relations team, and I'm here today to talk a bit with one of our GDE experts about AI dev tools. So, let's start. Can you introduce yourself and tell us a bit about what you do and how you've been using these products a bit? Hi Azar, thank you. Thank you for having me here. My name is Tomek Wierzchowski. I'm from Poland, from Bydgoszcz. I work as a staff ML engineer at Deep Sense AI, and I'm also Google developer expert for Cloud AI. Awesome. So, I'm...
