§1–§7
Rangeholder
Know which phone numbers will move, before you sign
Rangeholder is software that answers one question for people moving a contact center to a new provider: which of your phone numbers will actually come with you. Most are heads of support with an unsigned contract, a planned go-live date and up to 4,000 numbers. Their new provider checks a port only after signature, so a late answer means paying two providers.
Rangeholder keeps one row per phone number you own, checked against published UK and US numbering data, with one of four answers. It will move, with the rule named. It will move under a stated condition. It won’t move this way, with the alternative. Or no answer yet, with every date the software asked the provider in writing, once your porting lead approved the request. No answer ever comes from a model.
Machine learning at Rangeholder does one job: reconciling three customer files that never agree about the same numbers. One number can be a billing line, a label typed years ago and bare digits. A dictionary of known bill formats settles the well-formed majority. For the rest, a fine-tuned classifier weighs the label with the number’s 90-day usage and suggests live, dormant, fax or modem, or outbound-only. An embedding model suggests which descriptions across files are one line.
Rangeholder’s review queue is where a person decides every status a model suggests. A suggestion appears only above a fixed score and applies only when your operations manager confirms it. Below it, the row stays unclear and the list can’t be locked as a cutover baseline until someone decides. For 300 numbers the review takes two to three hours.
Every status confirmed in Rangeholder’s review queue becomes a labeled pair in a training set the company owns, shared across the platform. Each pair is a text label and usage profile matched to a person’s choice, with no calling numbers or call records. The classifier is trained on it.
Onboarding a site is Rangeholder’s heaviest job: every number needs a first answer within two working days. Millions of call records become a usage profile per number before both models run. Call records are your callers’ personal data, so no outside AI service sees them. Rangeholder is hosted on cloud infrastructure in Stockholm, and nothing runs anywhere else. The models use GPU instances that stop between jobs; the classifier moves to reserved GPU capacity there in Q2 2027, the matching model in Q4 2027.
Rangeholder AB was founded in Göteborg in November 2023 by a former support leader and two people from the carrier side. The first site’s list is built and delivered before any payment, so our heaviest compute runs before a new customer has paid.