Rent out a small AI that answers one yes-or-no question

Who pays

Developers and small software companies that need that one judgment at volume and don't want to train their own model. They would pay per call through a metered API, or pay a monthly plan that includes a set number of calls. This pricing is our proposal. The sources show training costs and accuracy gains, not revenue from anyone reselling a model this way.

How to build it

1. Define one narrow yes-or-no task

   Choose one narrow, repeated judgment, such as prompt-injection detection or fake-review detection, and write down exactly what counts as yes and no.

2. Train your own small model

   Install the Kev fine-tune skill with npx skills add jaredpalmer/kev@kev-finetune and run it in Claude Code to generate synthetic data and fine-tune on Modal. If you want to start for free, use the Unsloth Colab notebook instead.

3. Prove it beats general models

   Test the fine-tuned model on held-out examples against a general decision model, and publish the comparison only if your model actually wins.

4. Put the model online

   Deploy the model with the kev-deploy skill to a Modal endpoint, or put a Cloudflare Workers front end in front of it.

5. Issue keys and bill per use

   Issue API keys and log every call in Supabase, then connect those usage logs to Stripe metered billing for per-call or monthly-allowance plans.

6. Publish docs and pricing page

   Put a docs and landing page on Vercel that shows the task, the test results, the pricing, and a sample API call.

Sources on X

• jaredpalmer · Oct 8, 2026

  Jared Palmer, the author of Kev, shares a write-up of a $200 Kev-4B fine-tune that the poster says beats other decision models. The official kev-finetune skill finds decision-model callsites or interviews you, generates synthetic data, and fine-tunes on Modal. A sister skill, kev-deploy, deploys the result to Modal endpoints.

• jaredpalmer · Oct 2, 2026

  Jared Palmer claims that, depending on the use case, a 4B or even 0.8B Kev model can beat Jev for $10-30 with a Modal account. That figure is the poster's own claim.

• Sumanth_077 · Oct 11, 2026

  As of 2026-10-11, Unsloth was reported to have released a free Colab notebook. It fine-tunes Qwen3.5 4B with LoRA into a decision model and calibrates its output probabilities on held-out data.

• stas_sorokin_ · Oct 9, 2026

  In the poster's own test, a 31.8 MB open model was fine-tuned in 8 minutes on a MacBook using 1,386 public labelled texts. It beat Jev at prompt-injection detection, with F1 0.996 vs 0.956, at 33 ms per text and offline. The untrained base model scored the same 0.956 as Jev, so the poster attributes the whole gap to fine-tuning.

• clintsharp · Oct 10, 2026

  Clint Sharp of Cribl announced cribl-decision-1.0, a Jev-style model focused on telemetry. The poster expects many more decision models fine-tuned for specific domains. This is an enterprise example of the pattern, not a small-operator business.