Bringing AI to Everyday Products

There's a gap between what AI can do in a research demo and what it can do in a product used by 100 million people. Bridging that gap was the challenge with Paint Co-Creator — integrating DALL-E image generation directly into Microsoft Paint.

The Promise and the Problem

Generative AI is magical in demos. Type a prompt, get an image. But shipping that to millions of users introduces problems that don't exist in a Jupyter notebook:

Latency expectations — users expect Paint to feel instant, not "wait 15 seconds"
Content safety — you can't ship a product that generates harmful content
Cost at scale — every generation costs real money; multiply by 100M users
Discoverability — most Paint users aren't AI enthusiasts; the feature needs to feel natural

Architecture for AI at Scale

We built a pipeline that balanced quality, safety, and performance:

User Prompt → Content Filter → DALL-E API → Result Filter → UI
                  ↓                              ↓
            Block + Educate              Cache + Optimize

Key architectural decisions:

Credit-Based Access

Instead of unlimited generations, we introduced a credit system. This served multiple purposes — cost management, abuse prevention, and setting user expectations about the computational cost of AI.

Streaming UI

Rather than showing a spinner for 10+ seconds, we designed the UI to show progressive feedback. The user sees their prompt being processed, then a low-res preview, then the final image. Perceived latency dropped dramatically.

Safety as a Feature

Content safety wasn't a filter we bolted on — it was a core product feature. We built a multi-layered approach:

Input filtering — catch problematic prompts before they hit the model
Output filtering — analyze generated images before displaying
User education — explain why something was blocked, not just that it was

What Users Actually Want

The most surprising insight: users didn't want the most powerful AI features. They wanted simple, fun, and fast. The features that got the most engagement were:

"Make it better" — one-click enhancement of hand-drawn sketches
Style transfer — "make this look like a watercolor"
Background generation — "put my drawing in a forest"

The complex prompt-based generation? Used mostly by power users. The simple, one-tap features? Used by everyone.

Lessons for AI Product Engineers

After shipping AI to 100M+ users, here's what I'd tell any engineer building AI-powered features:

Latency is a feature. If your AI feature takes more than 3 seconds, you've already lost casual users. Invest heavily in perceived and actual performance.
Safety is not optional. Build it in from day one. Retrofitting safety into a shipped AI product is exponentially harder.
Design for the 90%, not the 10%. AI enthusiasts will find and use power features. But the real impact comes from making AI invisible and helpful for everyone else.
Measure what matters. Generation count is a vanity metric. Measure whether users keep the generated content, share it, or build on it.
Embrace constraints. Credits, rate limits, and content filters aren't limitations — they're design tools that shape a better product.

Looking Forward

We're at the beginning of AI becoming a standard part of every creative tool. The challenge isn't the AI itself — it's the product engineering around it. How do you make it fast, safe, useful, and delightful?

That's the work I find most exciting: not building AI for AI's sake, but making technology that helps real people create things they couldn't before.