AdGrid

Inspiration

We wanted to find a creative way to market our startup AgentGrid.sh. Giving our agents the ability to post a billboard ad right infront of the a16z office seemed like a good idea :).

What it does

AdGrid lets an AI agent take any company's website and autonomously turn it into a physical ad campaign, generating brand-aware creative (colors, logo, QR code pulled automatically), rendering it as static or video DOOH ads, previewing it as photoreal street mockups, and purchasing real screen time on San Francisco billboards through the CAASie API.

How we built it

We built a TypeScript agent on Guild.ai that chains a Remotion-based creative engine (brand extraction ‚Üí static/video ads at any dimensions), Sora 2 and nano-banana/gpt-image-1 via fal and OpenAI for video and photoreal in-situ mockups, and CAASie's API for booking actual billboard inventory.

Challenges we ran into

Physical-world platforms have physical-world constraints — CAASie's human artwork-approval loop and strict specs (like rejecting videos with audio tracks) forced us to engineer for compliance and approval latency, not just generation quality.

Accomplishments that we're proud of

An end-to-end run that goes from a URL to an approved-spec ad live on a real Market Street screen, plus shipping our creative engine as a standalone open-source skill anyone can use.

What we learned

Image models are great art directors but bad typographers — headlines and brand colors reproduce faithfully while URLs and QR codes drift — so pixel-exact compositing and generative realism have to be combined deliberately, not chosen between.

What's next for AdGrid

A marketplace where agents bid on ad inventory autonomously with on-chain payment rails (x402), campaign performance measured by QR scans, and creative that re-optimizes itself between flights.

Built With

• amazon-web-services
• caasie
• claude
• fal.ai
• guild.ai
• javascript
• markdown
• openai
• python
• render