We had 6 weeks and vague requirements: "we need chat and other AI features." What we built now powers multiple chat experiences and various AI features across web and React Native, with the architecture ready for Node.js SSR and Chrome extensions - from onboarding flows to business communications, email composition to PDF data extraction - with minimal changes along the way.
This talk reveals how we turned ambiguity into architecture by:

- Creating abstractions that work for any AI workflow (chat became just another use case)
- Building once, running everywhere: sharing code between web, mobile apps and other JS-powered platforms
- Generalizing from specific features to reusable capabilities
- Supporting streaming everywhere (because who doesn't need real-time AI?)

I'll share our journey from unclear requirements to production infrastructure, the architectural decisions that enabled this flexibility, and practical patterns you can apply to build AI features that scale across platforms and use cases.

[2 min] Introduction
Quick personal intro and setting expectations that this is about building infrastructure, not another ChatGPT wrapper.
[3 min] The Starting Point
The vague requirements we had ("chat and other stuff"), the platforms we needed to support, and why we couldn't just build a chat system.
[5 min] From Requirements to Architecture
The key insight that chat is just one type of AI workflow, the core abstractions that emerged, and how we avoided platform-specific traps.
[7 min] Implementation: Building for Speed and Flexibility
How capability-driven development let us ship working features weekly, the transport layer that unified different platform capabilities, and a real example of how the same abstraction handles chat messages and PDF extraction.
[6 min] Scaling Without Breaking
Generalizing from specific use cases like email to any text generation, how the chat layer enhances rather than limits, and supporting streaming everywhere with platform-specific optimizations.
[4 min] Results and Lessons
The various features built on top that we never expected, what required changes (very little), and the mistakes we made plus what saved us.
[2 min] Key Takeaways
When to abstract versus when to build specific, patterns for multi-platform AI infrastructure, and why treating chat as "just another feature" changes everything.
[1 min] Q&A Setup / Closing

Attendees will leave with concrete patterns for building AI infrastructure that works across platforms, practical abstraction strategies that enable flexibility without over-engineering, and the confidence to build diverse AI features beyond chat using a unified approach.