AI Transparency

Model Cards

Every AI model Syrka uses is documented here. We show what each model does, its limitations, and how we guard against bias.

/api/students/job-recommendationsdeepseek-chat

Recommend 6 real job roles based on student skills, sector-agnostic

Inputs
  • Student skills array
  • Country
Outputs
  • Job title
  • Match %
  • Salary range
  • Search URLs
temp: 0.5max_tokens: 2000
Limitations: May hallucinate salary ranges in smaller markets. Does not verify employer existence.
Bias Controls: Sector-agnostic prompt — no national vision weighting. Pure skill-match scoring.
Human-in-Loop: None — recommendations shown directly to student.
Data Retention: Request/response metadata logged to ai_audit_log. No PII stored.
/api/students/evaluate-offerdeepseek-chat

10-dimension weighted evaluation of a job offer

Inputs
  • Job title
  • Company
  • Salary offered
  • Student skills
  • Country
Outputs
  • Dimension scores
  • Overall weighted score
  • Grade
  • Negotiation points
temp: 0.4max_tokens: 1500
Limitations: Salary benchmarks may lag behind real-time market. Company reputation scoring is approximate.
Bias Controls: Weighted scoring with transparent dimension weights shown to student.
Human-in-Loop: Student decides — evaluation is advisory only.
Data Retention: Overall score and grade logged. Full evaluation not persisted.
/api/students/generate-cv-briefdeepseek-chat

Generate tailored CV brief with headline, ATS keywords, and cover letter opening

Inputs
  • Job title
  • Company
  • Description
  • Student skills
  • Country
Outputs
  • Headline
  • Skills to highlight/downplay
  • ATS keywords
  • Cover letter opening
temp: 0.5max_tokens: 1200
Limitations: Cover letter tone may not match all cultures. ATS keyword relevance is model-inferred.
Bias Controls: Vision alignment statement contextualised to student country.
Human-in-Loop: Student edits CV before submitting — brief is a starting point.
Data Retention: Headline logged for audit. Full brief not persisted.
/api/students/assess-ai-usagedeepseek-chat

Evaluate student submissions for AI collaboration sophistication

Inputs
  • Submission text (max 3000 chars)
  • Assignment brief
  • Student skills
Outputs
  • 5 dimension scores
  • Overall AI literacy score
  • Grade
  • Feedback
temp: 0.4max_tokens: 1500
Limitations: Cannot verify claimed AI usage. Assessment is inference-based from text patterns.
Bias Controls: Measures AI collaboration skill, NOT penalises AI use. Epistemic transparency is scored.
Human-in-Loop: Faculty reviews assessment before it affects grades.
Data Retention: Overall score and grade logged. Submission text not stored.
/api/students/adaptive-pathdeepseek-chat

Generate personalised 90-day learning path aligned with national vision

Inputs
  • Skills
  • Completed modules
  • Time available
  • Target role
  • Country
Outputs
  • Week ranges with actions
  • Velocity assessment
  • Bottleneck
  • Shortcut
temp: 0.5max_tokens: 2000
Limitations: Resource links may be stale. Time estimates are approximate.
Bias Controls: Velocity assessment uses neutral ahead/on_track/behind framework.
Human-in-Loop: Student chooses whether to follow recommendations.
Data Retention: Velocity and bottleneck logged. Full path not persisted.
/api/students/outcomesdeepseek-chat

Generate learning signal from application outcome data

Inputs
  • Job title
  • Company
  • Status
  • Rejection details
  • Skills gaps
Outputs
  • Priority skill to learn
  • Resource recommendation
  • Trajectory adjustment
temp: 0.4max_tokens: 800
Limitations: Learning signal quality depends on student-reported rejection details.
Bias Controls: Focuses on actionable skill gaps, not employer blame.
Human-in-Loop: Student logs outcomes voluntarily.
Data Retention: Priority skill logged. Full signal stored with outcome.
/api/university/evolve-curriculumdeepseek-chat

Generate updated reading recommendations with provenance data

Inputs
  • Course name
  • Course code
  • Description
Outputs
  • 3 recommendations with sources
  • Freshness score
  • ESCO skill codes
temp: 0.6max_tokens: 1000
Limitations: Sources are model-generated — may contain hallucinated DOIs. Faculty verification required.
Bias Controls: Freshness scoring penalises stale sources. Provenance verification flag defaults to false.
Human-in-Loop: Faculty must approve before recommendations go live.
Data Retention: Full recommendations and sources stored in curriculum_evolution_log.
/api/students/orchestration-scorelocal-calculation

Calculate AI orchestration readiness score from skills portfolio

Inputs
  • Skills array
  • AI assessment score
Outputs
  • Weighted score
  • Breakdown by 4 dimensions
  • Level
  • Next milestone
Limitations: Static skill lists may miss emerging AI tools. Score is formulaic, not adaptive.
Bias Controls: Transparent weight breakdown (30% AI skills, 25% breadth, 30% assessment, 15% vision).
Human-in-Loop: None — deterministic calculation.
Data Retention: Score and level logged to audit.
View Audit Trailarrow_forward