The model proposes, a rule ranks, a person confirms
The model proposes up to three scored methodology matches. A ranker with no model calls picks practitioners. Staff confirm every match before a plan runs.
- Supabase
- Custom
The problem
An intake has to land on the right methodology and the right practitioners. That is a judgment call, which makes it tempting to hand to a model, and nobody wants a model silently assigning a client.
What we built
The model gets the intake and the list of methodologies and returns strict JSON: up to three matches, each with a confidence score between 0 and 1 and a one-line reason, anything under 0.3 omitted. If there is only one candidate the model is skipped and the score is 1.0. If the output does not parse, the list is empty and staff choose by hand.
Inside the confirmed methodology, a ranker with no model calls picks the practitioners: a base weight, plus five points per tag shared with the intake, two per tag shared with the methodology's conditions, and one if research is already on file. Top three.
The match page preselects the top candidate and shows the score as a badge, High at 0.8 and above, Medium below. Nothing happens until a person clicks Confirm and Generate. "Choose a different methodology" opens the full list, zero candidates means choose manually, and a matching failure never blocks the intake from being created. The API also computes a 0.6 threshold; today it is informational, and every match is confirmed by a person.
The smallest model tier reads the intake (concerns, diagnoses, goals, duration) and proposes up to three methodology matches as strict JSON with a confidence score and a one-line reason. Anything under 0.3 is dropped. With a single candidate the model is skipped and the score is 1.0.
A practitioner ranker with no model calls scores practitioners inside the confirmed methodology by weighted tag overlap and picks the top three. The confidence score is shown to staff as a High or Medium badge; it never advances a match on its own. Staff must click Confirm before a plan is generated, can pick a different methodology from the full list, and a matching failure never blocks the intake.
Outcome
None published yet. What we can show: the ranker makes no model calls, the confidence score is displayed to the person confirming and never decides alone, and no plan is generated until someone has clicked Confirm on the match.
Who this fits
- Lead routing and ticket triage where the model should propose and a person should decide
- Services and marketplaces that match a request to a specialist
- Anyone who wants a confidence score on screen without letting it act alone