Built to a standard, node by node.
Fursah is described in the vocabulary of ITU-T Y.3172 clause 8.1, self-assessed against the 13 dimensions of the ITU AI Ready Report 2.0, and honest about what it does not cover. Every node below names the source file that implements it, so the claim can be checked rather than taken.
ITU-T Y.3172 (06/2019), Architectural framework for machine learning in future networks including IMT-2020, clause 8.1
The ML pipeline, in the standard's own terms
Clause 8.1 defines the pipeline as seven nodes. The left column of each row is what the Recommendation says the node is for; the right is what Fursah runs there. Keeping those two apart matters: describing our implementation in place of the standard's function would be a different taxonomy wearing the same identifiers.
Source
Supplies the data used as input to the ML pipeline.
Governed by PDPL · NDMO data classification
Evidence documents, employer role requirements, university offerings
prisma/schema.prismasrc/lib/documents.tssrc/actions/employer.ts
Collector
Collects data from one or more source nodes.
Governed by NCA ECC & CCC · DGA interoperability
Role-scoped ingestion through APIs, identity federation and institutional integration
src/app/api/src/lib/session.tssrc/lib/r2.ts
Preprocessor
Cleans, aggregates and otherwise prepares collected data before it reaches the model.
Governed by PDPL data minimisation · DPIA
Extraction, normalisation to the skill taxonomy, consent enforcement
src/lib/evidence-ai.tssrc/lib/careerTracks.tssrc/actions/documents.ts
Model
Hosts the machine learning models that produce the pipeline's output.
Governed by SDAIA AI Ethics · ISO/IEC 42001 & 23894
Deterministic scoring engine + grounded language model
src/lib/intelligence/readiness.tssrc/lib/ai.tssrc/lib/assistant/llm.ts
Policy
Carries the policies that constrain how the pipeline may operate.
Governed by SDAIA human oversight · PDPL rights
Consent rules, review thresholds, suppression floor, and the human override that binds M
src/actions/governance.tssrc/lib/cohort.tssrc/lib/policies.ts
Distributor
Distributes the model's output results to their destinations.
Governed by National Data Governance · cloud hosting rules
Releases each result only to the role authorised to receive it; aggregates suppressed below 5 students
src/lib/intelligence/ecosystem.tssrc/lib/cohort.tssrc/lib/assistant/context.ts
Sink
Receives the distributed output and acts on it.
Governed by DGA accessibility (WCAG 2.1 AA) · Arabic-first
Student, employer, university and policy interfaces
src/app/student/src/app/employer/src/app/university/src/app/workforce-intelligence/
The M node is split deliberately. The deterministic engine produces every score that affects a person; the language model only reads documents and explains results already produced. Human review carried at the P node overrides any output of M, and the override is logged. Read Y.3172 ↗
Two nodes the pipeline definition does not contain
Fursah runs two components that clause 8.1 does not define. They are listed separately rather than folded into the seven, because presenting a non-8.1 node as an 8.1 node is exactly the error this page exists to avoid.
Governance scenarios are evaluated against safeguards before a control is activated. The human decision, including any override, is recorded.
Every scoring surface stamps its model version onto the audit trail, so a result can be traced to the ruleset that produced it and that ruleset can be rolled back.
Self-assessment against the 13 dimensions
The report defines its dimensions from Plugfest projects, so one application is not expected to cover every dimension. Each row is marked addressed, partial, or out of scope. Dimension 9 is most relevant because it identifies skills gap analysis as a desired output.
The skill taxonomy is a shared reference that employers, universities and students all write against, which is the precondition for exchange. No marketplace or monetisation layer exists in the prototype.
src/lib/careerTracks.ts
Fursah generates no tradeable content. The language model reads documents and explains results; it produces no dataset or model asset intended for reuse or exchange.
The platform compares higher education with labour demand. It publishes the coverage gap between them as one figure.
src/lib/intelligence/ecosystem.ts
Built for the Saudi context rather than localised into it: the taxonomy, the evidence types, the Arabic interface layer, and the governance mapping to PDPL, SDAIA, NDMO and NCA instruments are all regional inputs, not translations of a foreign design.
src/lib/i18n/, src/lib/policies.ts
AI supports four defined points in the education-to-employment workflow: evidence extraction, readiness scoring, role matching, and curriculum alignment. Each point has a named input, a named output, and a later human decision.
src/lib/intelligence/
Arabic runs as a full runtime layer across every portal rather than a separate site, targets WCAG 2.1 AA, and the role-scoped assistant provides a conversational route to the same figures the dashboards show. The accessibility conformance claim is internal review, not an independent audit.
src/lib/i18n/translate.ts, src/components/FursahAssistant.tsx
The three stakeholder groups the report names are the platform's three portals, and the intelligence layer is the coordination mechanism between them. Alignment to the Human Capability Development Program is stated against specific commitments.
src/lib/nationalImpact.ts
Every extraction is a proposal a human accepts or rejects, and the rejection is retained. Students may dismiss a suggested career direction, and appeals against any automated result route to a named reviewer whose decision supersedes the model.
src/actions/documents.ts, src/actions/governance.ts
This is the platform's primary output. Fursah computes the skills gap at three resolutions: per student against a target role, per institution against employer demand, and per ecosystem as the set of requested skills no university offering covers.
src/lib/intelligence/readiness.ts, src/app/workforce-intelligence/page.tsx
The governance sandbox lets an operator state a proposed control, see which safeguards it breaches, and record the human decision. The workforce-intelligence surface is the evidence base a policymaker would review between statistical releases.
src/app/admin/governance/page.tsx
Fursah collects no gender, nationality, age or GPA field, so none can enter a ranking. Assessment is against published criteria identical for every institution, which is the mechanism by which a student from a less prestigious university is scored on evidence rather than on provenance.
prisma/schema.prisma, docs/DPIA.md
Career tracks carry per-skill weights, and universities set their own offerings, so priorities are expressible at institution level. There is no mechanism yet for a region or sector to set its own weighting over the national taxonomy.
src/lib/careerTracks.ts
The Y.3172 nodes are identified and mapped above. Infrastructure readiness is a national measure rather than an application measure. The prototype's hosting remains a declared implementation gap.
src/lib/standards.ts
Policy gaps this project ran into
These implementation gaps follow the three categories in the AI Ready Report. Each gap records a constraint that Fursah cannot resolve alone and states what is needed to close it.
International standards
Matching a course outcome to an employer requirement requires both to name the same skill. No national or international taxonomy is authoritative here, so Fursah carries its own seeded reference table.
This is a data-harmonisation gap of the kind chapter 4 names directly. Every platform in this category invents its own taxonomy, which makes results incomparable between platforms and prevents an institution from carrying its mapping to another system.
A standardised, versioned skill taxonomy with a defined extension mechanism, so that a curriculum mapping made once is portable and two platforms' readiness figures mean the same thing.
SDAIA, Ministry of Education, ETEC and sector skills councils
Two institutions or platforms exchange course, credential or vacancy requirements.
At least 95% of exchanged skills resolve to a versioned national identifier; extensions carry an owner and review date.
Twice yearly, with emergency additions for regulated occupations
A human reviewer approves an uploaded certificate and it becomes verified evidence inside Fursah. That verification cannot leave the platform: another system must re-verify from scratch.
There is no standard representation for 'this evidence was checked by a named party under a stated procedure' that a receiving system can evaluate. Verification effort is therefore duplicated at every boundary, which is the cost that keeps credential checking manual.
A verifiable-credential profile for skills evidence that carries the verifying party, the procedure applied and its date, so a receiving system can decide whether to accept it rather than repeat it.
Ministry of Education, ETEC and participating credential issuers
A verified skill or certificate is shared outside the system that reviewed it.
Every exported verification carries issuer, reviewer, method, date, status and revocation reference; receiving acceptance is auditable.
At issuance, revocation and annual trust-list review
National policy
Fursah deliberately collects no gender, nationality, age or GPA, so no protected characteristic can enter a score. The same decision makes disparate-impact testing impossible: there is no attribute to disaggregate outcomes by.
Data minimisation and demonstrable non-discrimination pull in opposite directions, and no instrument we could find resolves which takes precedence for an employment-adjacent system. Proxies remain: institution, region and career interruption can each stand in for a protected class.
A lawful basis for holding protected attributes strictly for fairness auditing, held separately from the scoring path and accessible only to an auditor. Without it, every minimising system in this category is structurally unauditable.
SDAIA/NDMO with the Ministry of Human Resources and Social Development
An employment-adjacent scoring system enters a controlled pilot or records 100 consequential recommendations.
Complete separation between audit attributes and scoring inputs, quarterly proxy review, and documented remediation for any material disparity.
Quarterly during pilots; annually after approval
Application hosting, object storage and model inference all currently run outside the Kingdom. The DPIA records this as risk R5 and marks it blocking for production.
PDPL transfer conditions are clear about the obligation, but a small project has no accessible compliant inference option: the affordable model-serving platforms are all extraterritorial, and the in-Kingdom alternatives are procurement relationships rather than services one can sign up for.
A published tier of in-Kingdom inference reachable by research and prototype workloads, or a defined sandbox basis under which pre-production systems may use extraterritorial inference on non-production data with disclosure.
SDAIA, CST and approved in-Kingdom cloud providers
Before production personal data or identifiable evidence is sent to a model endpoint.
All production inference and evidence storage completed in an approved region, with processor and transfer records retained.
Before launch and after every hosting or model-provider change
Implementation
Fursah publishes no trend, growth or forecast figure anywhere, and the assistant is instructed to refuse trend questions, because the platform stores no historical series and none is available to check against.
Graduate and labour figures are published annually or quarterly by separate authorities on separate schedules and cuts. There is no joined education-to-employment outcome series at the resolution a matching system would need to know whether its recommendations worked.
A published graduate-outcomes series linking field of study to employment outcome at a suppressed but usable granularity. Without it, no platform in this category can demonstrate effect rather than activity.
GASTAT, Ministry of Education and the Human Capability Development Program
Annual graduate-outcomes publication and any platform effectiveness evaluation.
A documented, privacy-suppressed field-of-study-to-outcome series with stable definitions and at least three comparable periods.
Annual publication with quarterly quality review
Fursah ranks candidates and states the ranking is advisory. Nothing prevents an employer from screening by that ranking in practice, which would make it decisive without ever being labelled a decision.
The distinction between decision support and automated decision-making is stated in principle but has no operational test. A platform can satisfy every disclosure requirement while its output is used exactly as an automated decision.
An operational test for effective automation, such as pass-through rate, override rate, or a required minimum review. The obligation should depend on how the output is used.
Ministry of Human Resources and Social Development with SDAIA
A score, rank or recommendation is used to exclude, shortlist or materially prioritise a candidate.
Documented human review for every exclusion, monitored override and appeal rates, and suspension when review evidence is missing.
Monthly operational monitoring and quarterly governance review
Every document behind this page is public.
The knowledge base lists each instrument cited here with its publisher, a link to the original, and the file in this repository that depends on it.