The AI model call is the easy part of an AI product. Cost, privacy and testing are the hard parts.

15 minpractice video in, feedback for each goal out
4end-to-end browser test suites
0external calls in the browser tests
1command to deploy

The product

  1. A parent makes an improvement plan for a young athlete. The plan has phases and goals.
  2. The parent records a 15-minute practice video on a phone. The recording stops automatically.
  3. A multimodal AI model compares the video to the plan.
  4. The parent gets a score and specific feedback for each goal.

Our engineering decisions

  • Stream the video. The server streams the upload directly to a temporary file. It never keeps the full video in memory. Thus, small and cheap cloud containers are sufficient.
  • Better quality at the same AI cost. The AI provider charges for video length, not file size. Thus, we increased the video quality at no added AI cost. We also removed a second, unnecessary compression step.
  • Keep video private. The system is designed to delete all copies of the video after the analysis. This includes the copy at the AI provider. The deletion also occurs when the analysis fails.
  • Test without the cloud. One command starts the full system on a local computer. A test version of the AI model gives the same answer each time. Thus, the browser tests run with no external calls and no cost.
  • Deploy with one command. A script sets the permissions, deploys the services and checks the result.

What this means for you

An AI feature must have good economics, safe data and proof that it works. We engineer these properties together with the model integration.

Do you want to add AI to your product?

Tell us about your users. We reply with a plan, a cost estimate per use and a test strategy.

Start a project →