The information a graduate or international student needs most, which employers can actually sponsor them, exists only as a raw government register with more than 120,000 rows and no way to search it by what matters. The advice they get on their CV is generic: accepted or ignored, with no explanation, so nothing improves between applications. Careers teams have the same problem multiplied by a cohort, with no visibility of who is progressing and no way to intervene before it is too late.

The brief was to turn public data and a stack of adverts into a decision someone can act on today, and to give a careers team the same view across hundreds of people.

The register first: the full licensed sponsor list turned into a searchable product and kept current automatically, refreshed every day, so a new sponsor is visible the same day rather than in a quarterly spreadsheet. Then the CV: scored against a specific advert, with the weak sections rewritten rather than merely flagged, and the same scoring available inside the browser on the advert the user is already reading. Then interview practice as a real spoken conversation with structured feedback afterwards, and adaptive question practice that responds to performance. Live vacancies are matched against both the sponsor register and the person’s own CV. For institutions, cohort seat licensing with automatic enrolment and progress analytics for the careers team.

The platform runs on web, native iOS and Android apps, and a browser extension, which is three surfaces from one codebase and one set of practices.

Consumer AI is a running-cost problem before it is anything else, because usage is unpredictable and every request spends money. Cost attribution per feature and per user, ceilings enforced in real time, caching and batching wherever the workload allows, and routing routine work to economical models are not optimisations added later; they are what makes a consumer AI product survivable. Voice interview practice in particular taught us to design the cost of a conversation before designing the conversation.

The other lesson is discipline about data. A platform that reads CVs holds personal data at scale, so the lawful basis, retention and processing records are written down, the models used do not train on user content, and the same position is what we recommend to clients handling candidate or patient data.

If your project involves turning a large public or internal dataset into a product, scoring documents against criteria, spoken AI, or shipping the same product across web, mobile and a browser extension, this is the evidence that the team has done it and runs it. The institutional side is also the proof for our training-provider and education work.