How it works — Redeploy

Step 1 — Drop the roster, any way it comes

CSV or Excel. Danish headers, English headers, typos, abbreviations — it does not matter. The engine recognises Medarbejder-ID, Emp_ID and employee id as the same thing, and it will tell you which columns it mapped before anything runs. Nothing is guessed silently.

Step 2 — Hard constraints first

Before any scoring happens, the engine applies the rules that cannot be bent:

Capability family — a person from Clinical Development is not matched into Supply Chain. The family must be compatible.
Sub-family — within the family, the specialisation must fit (QA, Data Science, PV…).
Location — the role must be where the person can work, with sensible synonyms handled (Bagsværd, Bagsvaerd, Bagsværd, Denmark are the same place; remote and multi-location postings accept anyone).

People who clear the constraints are scored. People who do not are never force-fitted — they get an honest "no exact fit" with the closest option listed for a conversation.

Step 3 — The score: three components, all visible

For everyone who clears the constraints, the engine reads three things and combines them with fixed, published weights:

45% · Title overlap

What the role is called

The position's title tokens must show up in the person's current role. Process Engineer matches Senior Process Engineer — and procesingeniør matches too.

30% · Keyword profile

What the job actually involves

Every Novo capability family has a weighted keyword profile built from the live job universe. The person's own position text is scored against the family profile — not against a model's vibes.

25% · Description coverage

What the posting asks for

The person's skills are checked against the posting's responsibilities. The more of the person's profile the posting asks for, the higher the coverage.

The weights are fixed and published — that is the point. A works council or an auditor can read the rulebook in one sitting and replay any decision.

Step 4 — The written reason

Every row in the output carries its verdict and its why: which constraints passed, what scored where, and why the winner won. Three tiers come out of the rulebook:

Match — clears constraints, score above threshold. Recommended.
Close — below the match threshold, but the best available fit. Worth a conversation or a relocation.
None — no fit in today's universe. Recorded with the closest option, so nobody is left with silence.

The messy-data promise

Real exports are ugly. The engine was built on a 250-person noisy benchmark — typos, Danish/English mixes, half-empty cells — and still puts the correct role in the top three for 97.6% of people. It stems words (engineering ≈ engineer), repairs typos within a bounded edit distance (acountant → accountant), and translates common Danish HR vocabulary. All deterministic. All explainable.

Determinism is the feature

Run the same roster twice — today, next week, on a different machine — and the output is byte-identical. There is no sampling, no temperature, no mood. When a decision is challenged, you replay it, and it answers the same way. That is what makes a redeployment programme defensible in front of a union, a regulator, or a courtroom.

A rulebook, not a guess.