Obscure filtering
ATS ranking is a black box. A candidate cannot know why they were filtered out, and an employer cannot check whether the criteria are reasonable. There is no basis for appeal.
Explainable AI for
workforce readiness
One shared skill language connecting students, employers and universities with deterministic scores, bounded generative AI and human authority over every consequential decision.
7/7 Y.3172 nodes mapped
13/13 readiness dimensions assessed
17 authentic sources
Working prototype at fursah.org
Team Visionary
Shoug Alomran · Lolwah Alsaadoun · Renad Alsulaiman · Taleen Bin Nader
info@fursah.org · Riyadh, Saudi Arabia
The route from university to work is filtered by processes nobody can inspect
Employers sift CVs through keyword-driven applicant tracking. Universities receive no signal about whether their graduates are hired. Students receive a rejection with no reason and no way forward.
Credentials over skills
Screening leans on institutional reputation, GPA and job titles instead of measuring the capability the role actually requires.
Disconnected education
Curriculum review cycles run years behind labour-market change. Employer demand never reaches programme design as a usable signal.
Hidden discrimination
Historical hiring data carries past preference. Models trained on it discriminate by gender, region or institution unless explicitly checked.
Uncontrolled data
Evidence of a student's capability is fragmented across organisations, with no verification chain and no student control.
Fursah turns three disconnected views into one shared skill language
Employer requirements, evidence-backed student profiles and university offerings are normalised against one versioned taxonomy. From that single layer the platform produces a different, role-appropriate output for each participant.
- Students get a Career Readiness Score, named gaps and the next best action.
- Employers get job-related fit explanations they can defend and audit.
- Universities get privacy-suppressed demand signals tied to curriculum action.
- Administrators get evidence review, audit events, appeals and overrides.
Verified evidence
Skills, certifications, projects and experience. AI extraction proposes; a human approves before anything becomes trusted.
Deterministic scoring
Readiness, gap, alignment and matching calculated from published, versioned weights, never by a language model.
Reasoning card
Every consequential output shows its contributing factors, ruleset version and the evidence behind it.
Fairness by design
Gender, nationality, age and GPA are never collected and cannot enter readiness or matching.
System truth boundary
Every design decision in Fursah follows from one strict separation: generative AI understands and explains, deterministic rules decide, and humans hold authority.
Generative AI may
- Propose structured fields from uploaded evidence, including certification details, project technologies, experience responsibilities.
- Explain authorised platform facts conversationally, scoped to the viewer's role.
- Declare its model and grounding versions with each answer.
Generative AI may not
- Calculate a score, rank a person or alter a ranking.
- Verify evidence or change a declared career direction.
- Invent a metric or expose another role's data.
Humans retain
- Evidence approval and reviewer attribution.
- Hiring decisions, curriculum decisions, appeals and overrides.
- Acceptance of production risk before deployment.
Verified 21 August 2026 in the deployed prototype: the assistant returned a grounded explanation through Cloudflare Workers AI (Llama 3.1 8B), repeating the same component figures the dashboard displayed and disclosing its grounding versions.
The proof chain
Evidence is extracted as a proposal
Uploaded documents are validated, privately stored and analysed. Extracted fields carry their supporting text and a confidence level, and remain advisory.
A human verifies before trust
AI extraction alone never makes evidence verified. A reviewer approves, rejects or edits, and the decision is attributed and audited.
Published rules calculate
Readiness, gaps, career alignment and candidate–role fit are computed from deterministic, versioned weights that anyone may inspect.
A grounded model explains
The assistant restates the displayed factors in plain language. The human reads the explanation and makes the decision.
Why the order matters
Reversing any two steps produces the system Fursah exists to avoid: an unverified extraction that scores a person, or a language model that ranks them. The chain is the control.
Seven pipeline nodes, mapped to running components
Evidence, role requirements, university offerings
Role-scoped APIs, sessions, institutional ingestion
Extraction, taxonomy normalisation, consent
Deterministic scoring plus grounded language model
Consent, review thresholds, suppression, override
Role-authorised output; cohorts under five suppressed
Student, employer, university and policy interfaces
| Node | Fursah implementation | Governing instruments |
|---|---|---|
SRC |
Student profiles, skills, certifications, projects, applications and uploaded evidence; employer requirements and hiring feedback; university offerings | PDPL & Implementing Regulations; NDMO data classification; National Occupational Standards for Data & AI |
PP |
Validation, private storage, AI-assisted extraction, taxonomy normalisation before any calculation | PDPL minimisation and purpose limitation; SDAIA privacy-by-design; DPIA |
M |
Generative AI bounded to extraction and grounded assistance; readiness, matching and workforce intelligence from versioned rules | SDAIA AI Ethics Principles; ISO/IEC 42001 & 23894; EU AI Act Annex III(4) |
P |
Human verification, employer hiring authority, student control of direction, appeals and auditable overrides | PDPL data-subject rights incl. objection to solely automated processing; ETEC assessment guidance |
SINK |
Student workspace, employer portal, university dashboard, administrator governance, shared workforce intelligence | DGA accessibility guidelines, WCAG 2.1 AA; Arabic-first bilingual interface |
Full node-by-node traceability, including source files, is published at fursah.org/standards.
All 13 dimensions assessed honestly, not universally
0Addressed: implemented in the prototype
0Partial: enabling component exists
0Out of scope: no claim made
0Assessed with evidence or a stated limitation
Addressed
Partial
Out of scope
Dimension 9 is the central contribution: gaps computed at student, institution and ecosystem
resolution.
Six stakeholders, one platform, distinct expectations
| Stakeholder | Need | Key expectation |
|---|---|---|
| Students | Understand readiness and improve skills | Clear gaps, actions and the right to appeal |
| Employers | Identify relevant candidates | Skills-based, explainable matching |
| Universities | Align programmes with market demand | Aggregated workforce insights |
| Administrators | Control risk and governance | Audit, oversight and escalation |
| Technology team | Reliable and secure platform | Controlled data and AI services |
| Regulators | Responsible AI and data use | Compliance and evidence of controls |
Business need
High
Value opportunity
High
AI suitability
High, with controls
Governance risk
High
Approach
Governed decision support
From fragmented judgement to a governed capability
- Student evidence is fragmented across organisations.
- Screening criteria are difficult to understand.
- Employer demand is not connected to curriculum decisions.
- Historical hiring patterns introduce proxy or rule-based bias.
- Students receive no useful feedback after an outcome.
| Stage | Target capability |
|---|---|
| Evidence | Verified skills, projects, certifications and experience |
| AI extraction | Structured extraction; advisory until verified |
| Assessment | Transparent, versioned readiness and matching rules |
| Explanation | Clear factors behind every important output |
| Human review | Mandatory review, override and appeal where required |
| Analytics | Aggregated labour-market and curriculum insight |
| Governance | Continuous monitoring, audit and escalation |
Every gap carries a requirement and a priority
| Gap | Requirement | Priority |
|---|---|---|
| Opaque screening | Explainable factors, rule versioning and an audit trail | High |
| Credential-heavy assessment | Job-relevant skills and evidence | High |
| Fragmented evidence | Evidence verification and controlled sharing | High |
| Bias and proxy risk | Rule-level fairness review, proxy-risk assessment, corrective action | High |
| Automation bias | Human override and reviewer controls | High |
| Weak education feedback | Workforce-demand and curriculum analytics | Medium |
| Limited evidence types | Portfolio and practical evidence pathways | Medium |
Root causes trace one level deeper: fragmented evidence → incomplete view of capability; opaque screening → low trust and no ability to challenge; weak feedback → slow workforce alignment; historical bias → potentially unequal outcomes; unclear AI authority → over-reliance on AI output.
Each risk is answered by a named control
| Risk | Impact | Level | Control |
|---|---|---|---|
| Unexplainable ranking | Unfair or unchallengeable outcomes | High | Explainability plus audit trail |
| Data misuse | Privacy and regulatory exposure | High | Purpose, access and retention controls |
| Proxy or rule-based bias | Potential systematic disadvantage | High | Rule monitoring, proxy-risk review, human oversight |
| Automation bias | Weakened human judgement | High | Override plus reviewer justification |
| AI extraction error | Incorrect evidence | Medium | Human verification before trust |
| Commercial influence | Biased recommendations | Medium | Sponsored content separated and labelled |
| Evidence exclusion | Some capability is missed | Medium | Alternative evidence pathways |
The governance approach is designed against SDAIA AI ethics and adoption principles, PDPL requirements, applicable national cybersecurity controls, and AI management principles aligned with ISO/IEC 42001.
Sequenced by what must be true first
Establish a governed evidence layer
Outcome: trusted and traceable data.
Separate AI assistance from consequential scoring
Outcome: lower AI decision risk.
Formalise human oversight and appeals
Outcome: clear accountability.
Embed rule monitoring and proxy-risk review
Outcome: early detection of potentially harmful outcomes.
Connect employer demand to university analytics
Outcome: better curriculum alignment.
Strengthen privacy and security lifecycle controls
Outcome: lower compliance exposure.
Pilot targets, stated as criteria to be tested
These figures are projections for a Fursah pilot. They are not measurements, and they are not claimed as achieved impact. Each is a target to be evaluated during and after a governed pilot.
0–0%Reduction in recruiter screening time per vacancy
0–0%Reduction in days from vacancy to qualified shortlist
0–0%Improvement in gap closure before graduation
0–0%Improvement in match quality and conversion
| KPI | Target | Why it matters |
|---|---|---|
| Placement rate | +10–15% vs. baseline | Tests the workforce outcome |
| Explanation satisfaction | ≥85% | Tests explainability |
| Appeal resolution | ≥90% within SLA | Tests accountability |
| Override rate | Baseline set during pilot, measured continuously | Detects automation bias or weak rules |
| Rule and proxy-risk review coverage | 100% of high-impact rules | Tests governance readiness |
The response is pre-agreed, not decided under pressure
Commercial capture
Sponsored employers or courses must be labelled, excluded from scores and auditable. If separation fails, ranking is suspended under the Responsible AI policy.
Inferred-attribute advertising
Inferred readiness, disability or socioeconomic signals are prohibited from advertising and profiling. Sharing stops and the matter escalates under PDPL.
Potential disparate impact
Investigate rules, thresholds and proxies using separately governed audit data; document remediation; suspend where necessary.
Automation bias
Require written justification for exclusions, monitor unusually low override rates, and retrain reviewers to treat explanations as support, not authority.
The non-standard candidate
Accept portfolios and practical evidence; present a low score as incomplete evidence rather than a judgement of competence; provide appeal and override.
Small cohorts
Suppress groups below five, withhold statistics, and partition outputs to prevent complementary disclosure.
Six gaps met while building, each with an owner and a trigger
| Encountered gap | Accountable owner | Operational trigger | Success measure |
|---|---|---|---|
| Fairness audit data | SDAIA/NDMO + MHRSD | Controlled pilot, or 100 consequential recommendations | Separate audit attributes; quarterly proxy and disparity review |
| In-Kingdom inference | SDAIA + CST + approved cloud providers | Before identifiable production data reaches a model | All production inference and storage in an approved region |
| National skill taxonomy | SDAIA + MoE + ETEC + skills councils | Cross-institution skill exchange | 95% resolution to versioned national identifiers |
| Portable verified evidence | MoE + ETEC + credential issuers | Verified evidence leaves the reviewing system | Issuer, reviewer, method, date, status and revocation retained |
| Graduate outcome baseline | GASTAT + MoE + HCDP | Annual outcomes publication | Stable suppressed series across three comparable periods |
| Effective automation threshold | MHRSD + SDAIA | A score excludes or prioritises a candidate | Human review evidence; override and appeal monitoring; suspension if absent |
Policy position
Employment-adjacent AI should be governed by effective use, not product labels. If a ranking excludes candidates without documented review, it functions as an automated decision even when it is called advisory.
Seventeen sources, each publicly checkable and load-bearing
The inclusion rule is strict: an entry qualifies only if it can be verified by anyone and something in Fursah depends on it. The live register names the publisher, edition, language, official URL, what the source contains, what depends on it, and the source files it constrains.
| Representative source | Contribution to Fursah | Publisher |
|---|---|---|
| ITU-T Y.3172 (06/2019) | Pipeline vocabulary and clause 8.1 node mapping | itu.int |
| ITU AI Ready Report 2.0 (2026) | 13 dimensions and the chapter 4 gap taxonomy | aiforgood.itu.int |
| Saudi Personal Data Protection Law | Minimisation, rights, consent and transfer conditions | sdaia.gov.sa |
| SDAIA AI Ethics Principles | Fairness, explainability, accountability, human oversight | sdaia.gov.sa |
| National Occupational Standards for Data & AI | Basis of the Fursah skill taxonomy | sdaia.gov.sa |
| NCA ECC and CCC | Cybersecurity and cloud-control baseline | nca.gov.sa |
| DGA Digital Accessibility | WCAG 2.1 AA target and bilingual service design | dga.gov.sa |
| ISO/IEC 42001 and 23894 | AI management and risk-management structure | iso.org |
| EU AI Act, Annex III(4) | Employment as a high-risk category | artificialintelligenceact.eu |
Structured JSON and CSV exports carrying the full pipeline, all 13 dimension assessments and every policy recommendation are available from fursah.org/api/knowledge-base, so the contribution is machine-reviewable rather than a list of links in prose.
Grounded in national strategy and independent evidence
Vision 2030 · Human Capability Development Programme
Fursah supports stronger connections between higher-education outcomes and the evolving skills needs of the Saudi labour market, addressing the education–employment divide the HCDP names as a long-term objective.
Market efficiency and matching
OECD research indicates AI can support labour-market matching by improving how skills, qualifications and opportunity information is processed. Fursah's own matching performance must still be validated by pilot.
Productivity and efficiency
OECD findings indicate workers gain productivity from AI in workplace processes, supporting AI-assisted document interpretation and explanation, with screening efficiency to be measured in deployment.
Responsible AI and algorithmic bias
Transparent decisions, human oversight, auditability and proxy-risk review. AI-assisted analysis is separated from consequential calculation, which stays deterministic and version-controlled.
Not a mockup: a working platform you can inspect
Student workspace
Readiness score, adaptive roadmap, career exploration, job matching, evidence management, privacy controls and data requests.
Employer portal
Job posting with evidence-based requirements, candidate fit explanations and hiring intelligence.
University dashboard
Cohort readiness, workforce demand, curriculum alignment and measurable action plans.
Administrator governance
Evidence review, reviewer attribution, appeals, overrides, audit events, cohort suppression and the operational Y.3172 trace.
Implementation
- Next.js on Vercel, bilingual Arabic-first interface, WCAG 2.1 AA target
- Prisma with libSQL/Turso for application data
- Cloudflare R2 for private evidence storage
- Cloudflare Workers AI (Llama 3.1 8B) for grounded assistance
- Verification suites asserting seven pipeline identifiers, 13 ordered dimensions, explicit limitations, HTTPS sources and operational policy fields
Privacy suppression, assistant context boundaries, evidence approval rules, reviewer attribution and submission completeness all pass. Live behavioural assistant verification succeeded in the deployed prototype.
What Fursah does not claim
Demo accounts are synthetic. The following require governed external evaluation and cannot be self-certified by the team, so they are presented as validation work, not prototype achievements.
Placement impact
No improvement in placement rate or time-to-hire is claimed. Those are post-deployment outcomes needing real participants.
Independent fairness validation
Lawful fairness-audit data and expert validation of scoring weights remain external work.
Accessibility certification
WCAG 2.1 AA is a declared target. An independent accessibility audit has not been performed.
Production hosting approval
Production stays blocked until hosting regions, processor arrangements and cross-border transfers are institutionally approved.
Governed pilot outcomes
Every KPI on this deck is a success criterion to be tested, not a measurement already taken.
What is claimed
A running pipeline, a reconstructable score, a bounded assistant, an auditable human decision and reusable policy evidence.
One proof chain, end to end
| Step | Live evidence |
|---|---|
| 1 | Open fursah.org/judge-demo and select Abdullah Al-Ghamdi from the prepared accounts. |
| 2 | Reconstruct 36/100 from the five weighted readiness components. |
| 3 | Ask the assistant why the score is what it is; confirm it repeats displayed facts and declares its model and grounding versions. |
| 4 | Switch to Administrator; review AI extraction proposals, human verification and reviewer attribution. |
| 5 | Inspect appeals, overrides, cohort suppression and the operational Y.3172 trace in Governance. |
| 6 | Use the employer and university accounts to follow skill demand into curriculum action. |
Fursah does not ask you to trust a black box.
It asks you to inspect
one.
Inspect the pipeline. Reconstruct the score. Challenge the explanation. Follow the human decision. Reuse the policy evidence behind it.
Team Visionary: Shoug Alomran, Lolwah Alsaadoun, Renad Alsulaiman, Taleen Bin Nader
info@fursah.org
Riyadh, Saudi Arabia