CareerProof Job Search Agent

• 

  Why trust matters in job discovery

• 

  Reviewable discovery-to-approval workflow

• 

  Fictional judge path and passing tests

• 

  Build Week V3 delta and Codex evidence

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for CareerProof Job Search Agent

Inspiration

Job-search tools often optimize for application volume, while candidates need confidence: why is this role relevant, have I already acted on it, and exactly which version of my materials would be sent? CareerProof makes discovery helpful while keeping every external action permissioned.

What it does

CareerProof ranks fictional opportunities against an editable profile and produces transparent dimension scores, fit reasons, concerns, and hard filters. It remembers feedback, prevents duplicate queue entries, enforces a strict preparation state machine, and binds the exact reviewed artifact to the target and channel with SHA-256. Every manifest remains awaiting approval; the judge demo performs zero external actions.

How we built it

The public Build Week edition is a dependency-light Python 3.11 package with an inspectable CLI, JSON Schemas, deterministic hashing, local state, pytest coverage, fictional fixtures, and CI. Codex with GPT-5.6 helped implement the V3 workflow, diagnose a ranking issue, add the schemas and regression tests, audit the public snapshot, and write the judge path.

Challenges

The hardest design problem was separating helpful automation from authorization. A high score must not bypass a hard candidate constraint, refreshing the job pool must not duplicate in-progress work, and changing one byte of the final artifact must invalidate its approval identity.

Accomplishments

• Explainable scoring with user constraints that outrank baseline fit.
• Persistent feedback and deterministic queue identities.
• A state machine that rejects unsafe jumps.
• Hash-bound, idempotent submission preparation.
• A privacy-safe public snapshot with eight passing tests.
• A local judge path requiring no account or API key.

What we learned

Trustworthy personal agents need explicit state, provenance, and approval more than they need another autonomous apply button. Making the safe path inspectable also made it easier to test.

What's next

Add opt-in job-source adapters, human-reviewed platform question cards, signed receipts for approved adapters, and public benchmark evaluation for explanation quality.

Codex and GPT-5.6 evidence

GPT-5.6 was used through Codex for development and review; the runtime judge demo is deterministic Python. Devpost /feedback Session ID: 019f7137-6601-7813-9cff-48c83edf70b0.

Built With

• pytest