WorkRadar AI Job Agent

From a fixed job match to an explainable, candidate-controlled readiness plan.

Comment

Inspiration

A job-match score can show that an opportunity may fit, but it does not tell a candidate what to verify or do next. WorkRadar AI Job Agent was created to turn a fixed match into a clear, explainable, candidate-controlled readiness plan.

The goal is to help candidates move forward without allowing AI to change the original match score or make employment decisions for them.

What it does

The prototype starts with one of three synthetic WorkRadar job scenarios. Each scenario has a fixed match score: 92%, 88%, or 90%. These scores are treated as immutable inputs and are never recalculated by the AI.

The candidate chooses a personal priority, such as:

  • getting ready as quickly as possible;
  • reducing risk before applying;
  • improving long-term career readiness.

GPT-5.6 Sol then generates a readiness plan based on that selected priority. The plan separates:

  • confirmed facts;
  • items that still need verification;
  • recommended candidate actions;
  • explanations supported by approved synthetic evidence IDs.

Changing the candidate’s priority changes the readiness plan while leaving the original match score unchanged.

How we built it

Codex was used throughout the project to plan, build, iterate, review, safeguard, and test the experience.

The application uses a controlled server-side integration with GPT-5.6 Sol. Model input is limited to synthetic scenario data, the selected candidate priority, and an approved list of evidence IDs. The response is validated before it is displayed.

The API key is stored only as a protected server-side secret and is never exposed in the browser or committed to the repository. Model requests use store: false.

Safety and privacy

This competition prototype is completely isolated from WorkRadar production systems.

It does not contain or access:

  • real candidates or resumes;
  • real users or employers;
  • production databases;
  • WorkRadar’s private matching formula;
  • production authentication or infrastructure.

The AI does not contact employers, submit applications, decide eligibility, rank real people, or make employment decisions. The candidate remains in control of every next step.

Challenges we ran into

The main challenge was preserving a strict boundary between matching and readiness guidance. The AI needed to create useful, personalized plans without changing the original match or inventing unsupported facts.

We addressed this by keeping match scores outside the model’s control, restricting recommendations to approved synthetic evidence, validating outputs, and clearly separating confirmed information from open checks and candidate actions.

Another challenge was presenting explainability without overwhelming the user. The interface therefore focuses on a simple flow: fixed match, evidence review, candidate priority, readiness plan, and human decision.

Accomplishments that we're proud of

We successfully built and tested:

  • three synthetic job scenarios;
  • three candidate priorities;
  • fixed and immutable match scores;
  • priority-sensitive readiness plans;
  • evidence-bound recommendations;
  • explainability through synthetic evidence IDs;
  • server-side secret handling;
  • an isolated competition repository and private live demo.

What we learned

We learned that a useful job agent does not need to replace human judgment. It can provide more value by organizing evidence, identifying open questions, and explaining practical next steps while keeping the candidate in control.

We also learned that strong AI boundaries can improve the product experience. Clear limits make recommendations easier to understand, verify, and trust.

What's next for WorkRadar AI Job Agent

The next step is to evaluate the readiness workflow with broader synthetic test coverage and candidate feedback. Any future production integration would require separate privacy, security, fairness, accessibility, and legal reviews before real data could be used.

For this competition, the project remains a synthetic, isolated prototype rather than a production employment decision system.

Built With

Share this project:

Updates