How FitRank decides
Most AI staffing tools ask a chatbot to pick someone. FitRank splits the decision into parts: code does what code can, a small model answers fixed questions, and a person decides.
- 1
Rules in code
Availability, leave, location, time zone, cost band and clearance are checked in code. Anyone who fails gets a recorded reason, not a low score.
- 2
Facts in code
Skill coverage, level gap, years of experience, domain projects and skill recency are calculated exactly, never guessed by AI.
- 3
Typed decisions
The in-house model answers five fixed questions, returning a probability for every option. It cannot answer outside the list.
- 4
Checks and bands
Code cross-checks each answer against the facts. A contradiction, thin data or an unsure model caps the person at Review.
- 5
A person decides
The manager accepts or rejects. Nothing is ever assigned automatically, and every decision teaches the next version.
Five typed decisions
Each question has a fixed list of answers. The model returns a probability for every answer and can't answer outside the list.
| Decision | Possible answers | What it judges |
|---|---|---|
| Skill match | none, weak, partial, good, excellent | How well the person's skills, including closely related ones, cover what the task needs. |
| Level fit | under, right, over | Whether their seniority suits the task. |
| Domain relevance | yes, no | Whether they've worked in the task's domain before. |
| Delivery risk | low, medium, high | Signals such as recent gaps, stale skills or overload. |
| Overall fit | yes, no | The final judgement, given the other four answers and the facts. |
Why probabilities, not a ranked list
A probability tells you how sure the model is. FitRank calibrates them, so when it says 90%, it's right about 90% of the time. Your thresholds turn that number into a band: Shortlist, Review or Hidden. A low-confidence answer never reaches the shortlist on its own.
Code checks the model
A well-formed answer can still be wrong, so FitRank compares each answer with the facts it calculated. A strong skill match for someone with none of the must-have skills, or “right level” for someone two grades away, is flagged as a contradiction and sent to Review. So are thin profiles and answers the model is unsure about.
A model that belongs to you
The matching model is a small transformer trained by distilling an open-weight LLM on thousands of example decisions. It runs on an ordinary CPU, so there is no per-run AI bill and no data sent to a third party for scoring. A hosted LLM is used only to read free text, such as a task description or a resume, into fixed fields. It never ranks or judges people.
It learns from your managers
The final score combines facts from code with the model's five answers, and that combination is learned from your managers' accept and reject decisions. FitRank retrains overnight on new decisions and reports what changed, but an admin decides whether to switch. Every model version can be compared and rolled back.
Explanations that cite facts
Every explanation points to numbered facts, such as “F1: must-have skills 3 of 3”. An explanation that cites a fact that doesn't exist is rejected and never shown.
Be one of the first teams to staff with FitRank
We're taking on a small number of pilot companies. Pre-register for early access, or talk to us about an annual licence.