Career Fit

Career Fit is an explainable, evidence-first job-search planner for job seekers, career coaches, and labor-market researchers.

It answers three questions that keyword tools usually collapse into one:

Can I do this? Can I prove it? Should I apply now?

Career Fit translates one job description and one candidate profile into a requirement–evidence map. It distinguishes direct evidence, transferable evidence, proof gaps, foundation gaps, and hard application gates. It then turns the most important gaps into concrete proof or verification actions.

Career Fit is not an ATS score, a hiring-probability model, an employer ranking system, or a judgment of personal value.

Why this product exists

Job seekers face different problems that look identical to a keyword checker:

• A career switcher may have adjacent experience but no employer-facing translation.
• A PhD or nontraditional candidate may have strong research capability but weak industry proof.
• A mid-career applicant may be ready on skills but blocked by an unresolved work-authorization, license, degree, or experience requirement.
• Anyone can have a proof gap: the capability exists, but the resume does not make it reviewable.

Career Fit keeps these cases separate so a user leaves with a better next move rather than a discouraging binary label.

What it does

Given a job description and a candidate profile, the v0.5 product:

1. extracts dictionary-backed skill requirements and explicit hard gates, including background-check, license, education, work-authorization, and experience requirements;
2. detects conservative local negation so statements such as “no direct HR-data experience” are not counted as positive evidence;
3. maps candidate statements to auditable evidence objects;
4. distinguishes direct evidence, thin proof, transferable evidence, and missing evidence;
5. presents the extracted checklist before any score, lets the user add a missed soft requirement or hard eligibility gate, then calculates Evidence Fit, Capability Signal, Proof Signal, and Application Readiness only after user confirmation;
6. ranks proof, translation, bridge, foundation, and verification gaps;
7. proposes an expected proof artifact and an evidence prompt for each priority action;
8. shows a multidimensional Role Fingerprint and the skill bundles that appear together in the supplied posting;
9. accepts a local plain-text or Markdown resume import and lets users label structured evidence as work, research, portfolio, GitHub, course, certificate, or claim;
10. exposes an explicit profile-language route so non-English profiles are not silently treated as fully mapped; users can add translated or structured evidence before relying on a score;
11. keeps claim-only evidence distinct from reviewable proof and aggregates multiple evidence items without letting a weak extra claim erase stronger proof;
12. offers a resume-free guided intake when mapped evidence is sparse: choose a role requirement and describe one task, context, optional result, evidence type, duration, and recency.

It can also compare two or three target roles against the same candidate profile after candidate evidence has been reviewed. The comparison carries that reviewed evidence state forward, while each role still requires its own checklist confirmation before role-specific scores are shown.

Quick start

Requires Python 3.10 or newer.

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -e .

# Analyze one role locally
career-fit analyze --job-file examples\single_job\people_analytics_job.txt --candidate-file examples\single_job\candidate_profile.txt --evidence-file examples\single_job\evidence.json
# Add --review-file review.json to reveal the reviewed score

# Compare a small target-role portfolio
career-fit compare --roles-file examples\role_portfolio.json --candidate-file examples\single_job\candidate_profile.txt --evidence-file examples\single_job\evidence.json --review-file examples\single_job\candidate_evidence_review.json

# Launch the visual explorer
career-fit serve
# Open http://127.0.0.1:8766

If the console script is not on PATH, use python -m career_fit.cli for the same commands. The legacy skillbundle namespace and console script remain available as a compatibility layer.

Optional semantic review

The explorer can send the supplied job text, candidate text, and deterministic requirement map to an OpenAI-compatible chat-completions endpoint for a second, text-based review. The review is advisory: it cannot change the rule-based scores, requirement classifications, or application decision. Leave the variables unset to keep the explorer fully local and rule-based.

$env:CAREER_FIT_LLM_API_KEY = "your-key"
$env:CAREER_FIT_LLM_BASE_URL = "https://api.openai.com/v1"
$env:CAREER_FIT_LLM_MODEL = "your-model"
career-fit serve

Candidate text is sent only when the review button is used. A local OpenAI-compatible endpoint may be used without an API key.

Visual explorer

The local page is designed to feel useful to a first-time job seeker while keeping the labor-economics logic visible:

• three animated signal rings separate capability, proof, and application readiness;
• the scorecard shows preparation signals only after review; before that it shows requirements identified, reviewable evidence, and an explicit eligibility status;
• the review panel lets a user remove false requirements, change importance, add known or user-supplied missed skills, confirm hard constraints, and add structured evidence before recalculating;
• the guided intake panel gives a resume-free user a short task/context prompt tied to a real role requirement, then carries the user-declared example through the same review route;
• a local resume text importer helps a first-time job seeker start without sending a file to a hosted service;
• the requirement–evidence matrix makes every assessment inspectable;
• the profile chart shows the shape of evidence overlap across requirements;
• action cards identify a time horizon, effort estimate, expected proof artifact, and evidence prompt;
• the Role Fingerprint shows category-level mismatch without replacing named-skill evidence;
• skill-bundle cards turn co-occurring requirements into integrated proof-artifact ideas;
• plain-language interpretation cards explain what the metrics can and cannot mean.

The design rationale and evidence boundaries are summarized in the literature-to-product map.

The 50-profile end-to-end release audit, route decisions, and remaining guided-intake boundary are documented in the job-seeker scenario audit.

Optional occupation context

Career Fit is specific to the job posting you provide. An optional occupation-context panel can connect a confirmed standard occupation to AI Labor Atlas and show a separate Market Context card with wages, employment, openings, projected change, AI exposure, representative tasks, and descriptive adjacent occupations, alongside public worker comments. Start AI Labor Atlas on http://127.0.0.1:8765, set CAREER_FIT_ATLAS_URL to that address, then search and confirm the closest occupation. The title suggestions are evidence leads, not an automatic classification.

For a small set of common nonstandard titles, Atlas may return a curated candidate occupation family rather than one exact title. Career Fit shows the mapping note and requires confirmation from the job tasks; these candidates do not change job-fit scores.

Worker comments remain separate from the job-specific analysis. They do not change Evidence Fit, Application Readiness, eligibility, or any other Career Fit signal. Comments can be positive, negative, mixed, incomplete, subjective, outdated, or biased; they are not verified facts or representative of all workers. Source, date, scope, and link are preserved where available. See the Atlas worker review contract for the import boundary.

Interpreting the scores

For each soft requirement j, the engine reports:

Match_j = 0.35 coverage
         + 0.25 evidence strength
         + 0.20 proficiency signal
         + 0.10 recency
         + 0.10 evidence depth

Evidence Fit is the importance-weighted average of these matches, rescaled to 0–100. Must-have, strongly preferred, preferred, and inferred requirements receive weights 1.0, 0.7, 0.4, and 0.2.

Capability Signal gives descriptive weight to direct, thin, and transferable overlap. Transferable evidence is intentionally discounted and never presented as equivalent proof.

Proof Signal summarizes how concrete and reviewable the supplied evidence is. Structured work, research, portfolio, duration, and measurable-result fields are stronger than a bare keyword mention.

Application Readiness is a preparation-triage measure:

Readiness = 100 × (0.50 must-have match
                  + 0.30 proof signal
                  + 0.20 hard-gate signal)

An unresolved hard gate can produce a verify_before_applying status even when Evidence Fit is high. None of these measures is a hiring probability or a causal estimate.

The first response is provisional. The review panel must be used to confirm hard constraints, correct extracted requirements, set importance, and label evidence before relying on the plan. Education is compared by ordered level, experience floors require matching experience-area text, background checks remain separate verification gates, and ambiguous free-text gates remain unknown. If fewer than two requirements are mapped, the job text or candidate profile is too short, or no candidate evidence is identified, the response is marked insufficient_information and no fit or readiness score is presented. Even when enough text is available, scores remain hidden until role requirements are reviewed. Claim-only evidence is visible as a claim but receives much less proof weight than reviewable work or project evidence. Non-English or mixed-language profiles receive a visible coverage caveat and can proceed through user-labeled evidence or trusted translation; the English-first dictionary does not silently claim full language coverage. requirements_identified is a count, not a completeness score; eligibility_status: no_gate_detected means that no gate was found, not that eligibility was verified. A fully evidenced role still ends with a role-specific application-positioning action.

The single-role machine-readable output uses schema career_fit.v0.5; role comparison uses career_fit.compare.v0.3. Details are in docs/data-contract.md. The scoring choices are documented in docs/methodology.md.

Data and taxonomy

The demo uses a versioned English dictionary with a transparent seed layer and an O*NET 30.3 enrichment layer. The current local build exposes 426 entries: Essential Skills, Transferable Skills, Knowledge elements, and Software Skills marked Hot Technology or In Demand. Exact matching is conservative: ambiguous software names require nearby software context, while bare common-word aliases are excluded. It is a transparent baseline for evidence extraction, not a complete occupational ontology.

• config/seed_dictionary_en.json stores canonical skill IDs, aliases, and analytical category codes.
• config/onet_enrichment_en.json stores O*NET-derived labels, source element IDs, exact-label mapping methods, and the 10-category analytical mapping.
• config/taxonomy_10_ai.json stores the ten-category research layer and 45 pair combinations.
• tools/build_onet_enrichment.py regenerates the enrichment layer from the registered O*NET 30.3 text release.
• config/source_registry.json records public sources and redistribution notes for ESCO, O*NET, and SkillSpan.
• src/skillbundle/career.py stores explicit transfer crosswalks and evidence rules.
• src/skillbundle/career.py also creates the descriptive Role Fingerprint and posting-specific skill bundles.

Raw third-party data is not silently bundled. Check each source’s current license and attribution requirements before redistribution.

Privacy and fairness

The rule-based demo runs locally. Candidate text and structured evidence are sent to the local Python process serving the page; if optional semantic review is configured and the user clicks its button, the supplied job text, candidate text, and requirement map are also sent to the configured review endpoint. This repository does not include a hosted personal-data service.

Career Fit should help a person prepare, not automate exclusion. A missing mention is not proof of missing ability. The engine excludes conservatively detected negative statements, exposes uncertainty, avoids protected traits, and keeps hard constraints visible for human verification. See docs/privacy-and-fairness.md.

Repository layout

src/career_fit/       public Career Fit namespace and CLI
src/skillbundle/      compatibility namespace and transparent engine
config/               English seed dictionary, taxonomy, and source registry
examples/             reproducible job, candidate, and evidence inputs
docs/                 product, design, methodology, and data contracts
tests/                extraction, negation, constraints, and fit tests

Roadmap

• expand the reviewed multilingual skill dictionary and occupational crosswalks;
• add resume-to-evidence import with explicit user confirmation;
• evaluate a separate career-pathway graph only after defining its transition data and validation strategy;
• add calibrated validation only if an appropriate, consented outcome dataset becomes available;
• publish sensitivity checks for taxonomy, importance weights, transfer rules, and missing-evidence assumptions.

License

MIT. See LICENSE.