Independent prototype, 2026

RoleProof

A job recommendation should explain itself. RoleProof compares the listing with your preferences and shows the evidence, the mismatch, and the questions still open.

Role
Solo project. Application, decision policy, evaluations, and interface.
Stack
TypeScript, React, Instructor JS, Zod, Express, Vitest, Promptfoo
Links
Live demo, GitHub
Code
MIT. Built on the open-source Instructor JS extraction library.

01

The problem

A role can mention the right technologies and still leave the practical details unclear. Remote can mean a particular country. A compensation figure can include equity. A missing technology mention can be mistaken for a mismatch.

I built a small explanation layer for these distinctions. Each preference gets one of three answers: supported, contradicted, or unknown. An unknown keeps the role in view and gives the applicant a concrete question to resolve.

02

The implementation

  1. Extract the listing. Instructor JS and Zod produce structured location clauses, pay ranges, technologies, and exact quotations. The extraction prompt does not receive the candidate preferences.
  2. Check the evidence. Code checks that each quotation occurs in the source. A missing quotation withholds the conclusion for that dimension.
  3. Compare preferences. A deterministic policy evaluates the stated hiring scope, comparable base cash, and technology mentions. Conflicting or incomplete evidence stays unknown.
  4. Make it reviewable. Source references highlight the original passage. Changing preferences recomputes the explanation. JSON export includes the source, facts, and decision.

The live browser demo uses the same decision code with clearly labelled saved facts. Five fictional job descriptions can be explored without an account or API key.

03

Decisions that mattered

Uncertainty

A missing fact is not a rejection.

Location ambiguity, a salary range that crosses the desired minimum, and an absent technology mention each remain visible as unknown. The policy does not invent an answer to fill the gap.

Reuse

Extract once, compare again.

Separating extraction from preferences lets a listing be compared with several profiles without asking the model to reinterpret it on every change. Instructor JS supplies the structured extraction infrastructure; the application policy and interface are project code.

Evidence

A quotation is a review aid.

An exact quote check establishes that text exists, not that a model understood it correctly. The passage stays available for inspection, and the evaluation harness keeps extraction quality separate from policy correctness.

04

Verification and scope

35 regression and integration tests pass, including the real Instructor package against a local fake HTTP provider. All 23 author-labelled synthetic policy scenarios pass in the native runner and Promptfoo replay. GitHub CI verifies the project on Windows and Ubuntu with Node 22 and 24.

These results measure specified policy and integration behavior. No live model accuracy or improvement over a commercial matching system has been measured. The repository includes a runnable live comparison against a single-prompt baseline, with correctness, false exclusions, unsupported confirmations, token use, and latency reported separately.

The prototype covers hiring-location wording, base cash, and technology mentions. It does not determine legal work eligibility or assess experience, seniority, time-zone overlap, or the full job requirements.

Try the live demo

Next case studyEvalgate →