Where should it run?
Callozum reads a request, finds the private data, splits the work into parts and picks where each part runs.
Why was this customer charged twice, and what should we tell them?
source
Photos: MTA Construction & Development (CC BY 2.0) and public domain, via Wikimedia Commons.
A request of your own needs Callozum's models live, and they run on Callozum's Worker, not this site's. Type it in the live demo on the Callozum site ↗. It works the same way.
What
The demo from Callozum, the router I'm building. Give it a request and its context. It says which model each part needs, which data is private, and where each part should run: on the device, a local edge, a cloud edge or the cloud.
How
- model
- Callozum (Clef, Clairic, Qwen3, Llama 3.3)
- id
@cf/cloudflare/clef@cf/cloudflare/clairic@cf/qwen/qwen3-30b-a3b-fp8@cf/meta/llama-3.3-70b-instruct-fp8-fast- runs on
- Callozum's Worker on Cloudflare Workers AI; here, saved runs in your browser
- costs
- free here: the six examples replay saved runs
- Classify. Clef answers 28 fixed questions about the request: the kind of job, the private data, the rules, the size, the speed and the place.
- Plan. Plain rules turn the answers into a plan. Private data is removed on the device before anything leaves.
- Draft. A fast LLM (Qwen3 30B) splits the request into parts, writes the reasons, and drafts the step 1 request for Clef or Clairic.
- Run. Step 1 goes to Clef or Clairic. Step 2 goes to Llama 3.3 70B, and only when new text is needed.
Why
Most AI work does not need a large model in the cloud. Sending everything there is slow, costs more, and moves private data off site. A small model can decide where each part goes in a fraction of a second. The case is in Inference That Knows Where to Run.
Try
- Leaked API key: the secrets are found and removed on the device before the fix is written.
- Patient note: health data that must stay in Canada. Open "Sensitive data".
- Robot fleet: fleet data pulls the work to the cloud. Then run steps 1 and 2.
Limits
- The examples replay runs saved from Callozum's Worker in October 2026. This page does not call the models.
- For a request of your own, use the live demo on the Callozum site.
- Times and costs in the comparison are estimates from published medians and prices.
- The LLM draft can change from run to run. Clef's answers are probabilities.
The same idea on a live stream. Requests from a robot arm, smart glasses or a bank chatbot go to the first place that is allowed and fits the time budget. Change the network, the battery or the data rule and watch the parts move.
Live sketchestimates
- Requests
- 0
- Dropped on the device
- 0
- Stayed on site
- 0
- Went over the internet
- 0
- Answered by System 1
- 0
- Private data sent out
- 0
Per hourlive, real rates
In the sketch, changed frames show up far more often than in real life, so you can see them. The numbers in this box use real rates. Filled parts are System 1 (fast checks: code, Clef, Clairic). Outlined parts are System 2 (language models).
Write to me about the problem you want to fix.
- emailEmail contact@meirdick.com
- linkedinLinkedIn in/mdick
- githubGitHub meirdick
- substackSubstack aistrategyconsultant
- rssRSS /rss.xml