Career CoDesk

Source evidence, provisional AI interpretation, capacity and human decision controls.
Adviser decision queue for learner cases requiring review.
An adviser-approved weekly action with inspectable provenance.

Inspiration

Further Education careers teams support many learners with limited adviser capacity and fragmented information. We wanted to explore whether AI could reduce preparation work while keeping every important decision with a qualified careers adviser. At where I live, a good Further Education college , Warwickshire College Group（WCG) has more than 11,000 students while on the website only 5 career advisor.

What it does

Career CoDesk is a decision and AI supported workspace for FE careers advisers. It brings learner evidence, student records, available capacity and possible support routes into one reviewable workflow. The system keeps source evidence separate from provisional AI interpretation. An adviser must approve, amend or reject every proposal before it becomes an active weekly plan. We are not try to replace the current adviser , we just want to free they time and enegery to support more students or learners. Learner feedback can also reopen a case without deleting its history.

It is not a job board or automated job-matching service. It is designed for education careers operations.

How we built it

The product is a Django application using Python and SQLite, with synthetic learner records and deterministic evaluation scenarios.

We built it with Codex and GPT-5.6 using a workflow inspired by OpenAI's Symphony orchestration specification. GPT-5.6 Sol provided high-level guidance and independent review. GPT-5.6 Terra and Luna supported design, implementation and verification through Codex CLI. almost all in automatic run, but with human gate to review and make decisions. Each implementation task ran in an isolated workspace. Deterministic tests and human approval gates—not the model—decided whether work was complete.

Challenges and what we learned

The main challenge was to discuss, within a short time, with the AI the project’s rationale and compliance, data processing and privacy management, and the relevant regulatory policies deciding Another challenge is what AI should not control. Capacity calculations, activation and final learner-support decisions remain deterministic or human-owned.

What's next

Our next steps are meetings with FE college careers teams, further product development, stronger integration and governance controls, and accessible learner-support features for adviser-approved follow-up and feedback.