← All Episodes
EP19September 14, 202600:48:43

Google Analytics for the Real World

Steve Walsh
CTO, Smart Retail

Steve Walsh — CTO of Smart Retail, an edge-based computer vision platform bringing measurement to out-of-home and physical retail. Originally from Northern Ireland, now in Hong Kong. Runs small models on-device rather than in the cloud, so the raw video never leaves the premises and only aggregate metadata is uploaded. Met the show at the Snowball panel event.

Computer VisionEdge AIRetail

Chapters

  1. 0:00Intro
  2. 0:14What are we doing at Compass Studios?
  3. 0:45Meeting Steve at the Snowball event
  4. 1:28Smart Retail, and thanks for Overcast
  5. 3:07We built our own model
  6. 5:50From Northern Ireland to Hong Kong
  7. 7:44A Cantonese village or a Hakka village?
  8. 8:42What Smart Retail actually does
  9. 9:39Measuring advertising in MTR stations
  10. 11:26Proving a campaign actually worked
  11. 12:22Google Analytics for the real world
  12. 13:16How a real AI model gets built
  13. 15:13Running it on a low-power device
  14. 18:50Twenty metres at near-100% accuracy
  15. 19:48Shrinking the model
  16. 20:41Controlling the whole stack
  17. 21:34Everything runs on NVIDIA GPUs
  18. 23:29Calibration matters more than the model
  19. 25:22Side projects, and how he got here
  20. 27:08Using Codex for review
  21. 29:01Can AI actually help you ship?
  22. 30:54Staying motivated and on target
  23. 32:48LLMs packaging the whole thing
  24. 33:44Iterating on design in minutes
  25. 34:43It's just an API a model can talk to
  26. 37:21The new paradigm
  27. 39:15Requirements, then a PRD
  28. 42:02Deploying safely without breaking things
  29. 42:55Automating the test suite
  30. 43:54Most of our code is now LLM-generated
  31. 44:47Supervising the factory
  32. 46:41Building a better factory

About This Episode

Online, every impression is measured. Step outside and the whole industry goes blind. Steve Walsh is CTO of Smart Retail, an edge computer vision platform that brings web-style analytics to physical space — and he is careful to explain how it counts people without identifying them. The engineering is the interesting part: a 30-to-40-million-parameter model small enough to run on a standalone device with no connection, uploading only aggregates. Also: what four years of AI coding tools actually changed about how his team ships.

Watch on YouTube Listen on Spotify Listen on Apple Podcasts Listen on OvercastAll Episodes