← Overture

// status banner › v1 to v2

Rules vs. a model, live

On Overture, attorneys post referrals and other attorneys respond to them. The referral page showed the case details but almost nothing about what was happening with it: how many people had looked, whether anyone had responded, whether it was still worth answering.

I designed and built a fix: a stats row (views, responses, last response), a live "posted" counter, and a banner that picks a message from a set of rules, like "Just posted" or "A few replies are in." I opened the PR, went on parental leave, and while I was out the company refocused on its law firm business. It never shipped.

At the time I'd also pitched an AI summary for the banner. This page rebuilds the original feature on mock data and adds that second version, so you can compare them:

  • v1, rules: picks a fixed message based on thresholds
  • v2, Claude: writes a short summary from the same stats, plus the matter details and the profile of the attorney viewing it

The rules work, but they can only say what someone wrote a template for. A referral can be busy but going quiet, or be a strong fit for one attorney and a poor one for another. v2 can say those things.

v1 didn't get thrown away. It became v2's guardrail. The model also returns a category (new, trending, stalled and so on), and the server checks it against v1's rules before anything reaches the page. If they disagree, if the text includes a number that isn't in the input, or if it claims something about the viewer that isn't in their profile, you get v1's message instead.

What this isn't: it never ran in production, so there's no user data behind it. The next step would be shipping v1 and measuring response rates before adding the model.

Live demo, running on mock data. v2 is a real model call, made only when you ask.