Overview / AI Architecture System Design & AI Craft

AI Architecture & Career OS Integration

How Lumina Graph plugs into Talentbank Career OS as an API-first intelligence layer.

Data & Processing Flow
Inputs
📄 Resume / CV
🏆 Credentials
🎓 Activities
↓
⚡ Lumina AI Extraction Engine
LLM Parsing · Asian Talent Graph · Inferred Skills · Evidence Mapping
↓
Skills Graph Engine
Explicit + Inferred Skills · Confidence Scores · Taxonomy
Learning Velocity Engine
Skill acquisition speed · Recency weighting · Trajectory
Readiness Engine
ARS · Domain readiness · Gap analysis · Signal score
↓
🎯 Smart Matching Engine
Semantic Vector Search · Cosine Similarity · Predictive Fit Score · Explainable AI
↓
Candidate Dashboard
Living Portfolio · Scores · AI Insights · Learning Path
Employer Dashboard
Ranked Candidates · Explainable Match · Shortlist
University Dashboard
Cohort Readiness · Skill Gaps · Recommendations
↓
🔌 Talentbank Career OS Integration API
REST API · Headless Architecture · Plug-in as Module · No Platform Rebuild Required
Prototype Tech Stack
🎨
Frontend PHP 8.2+ · HTML5 · Tailwind CSS (CDN)
📊
Charts Chart.js 4.4 — Radar, Bar, Doughnut
⚙️
Logic Alpine.js / Vanilla JS — AI simulation
🗃️
Data PHP arrays (demo) → MySQL-ready schema
🤖
AI (Demo) Deterministic scoring engine (demo mode)
⚡
AI (Live) Claude / OpenAI / Gemini API (config.php)
☁️
Hosting Hostinger · PHP shared hosting · File upload
🔒
Auth Demo credentials (PHP session-ready)
Production Architecture (28-Day Build)
🅰
Frontend Angular 17+ (Talentbank Tier 1 Adoption)
⚙️
Backend NestJS · TypeScript · REST API + Swagger
🗄️
Database PostgreSQL + pgvector (semantic search)
🧠
AI Claude Sonnet / OpenAI embedding-3-small
📐
Vector DB pgvector · cosine similarity · 1536-dim
🔑
Auth Supabase Auth · Google OAuth · JWT
🚀
Deploy Vercel (frontend) · Railway (backend)
📋
Monitoring Audit logs · AI usage disclosure · PDPA
AI Craft — How It Works
🧠 Skills Extraction Engine
  • › Resume text → LLM structured JSON
  • › Explicit skills from stated experience
  • › Inferred skills via Asian Talent Graph
  • › Confidence scoring per skill
  • › Evidence mapping: source → skill
⚡ Learning Velocity Engine
  • › Time-series skill acquisition tracking
  • › Complexity progression measurement
  • › Recency decay weighting (λ=0.003)
  • › Credential and project completion speed
  • › Percentile ranking within cohort
🎯 Semantic Matching Engine
  • › JD → skills vector (1536-dim)
  • › Candidate profile → embedding vector
  • › Cosine similarity via pgvector
  • › Multi-factor Semantic Fit Score
  • › Natural language explanation (RAG)
AI Governance & Ethics
🔒
Data Minimisation
Only necessary data collected. PII scrubbed before AI processing.
🙈
Blind Screening Option
Employers can hide names, universities, demographics. Pure skills-based ranking.
🧑‍⚖️
Human-in-the-Loop
Lumina produces signals, not decisions. Final hiring call always with humans.
📢
AI Transparency
"AI-Generated" labels on all outputs. Model usage disclosed in audit logs.
⚖️
Bias Monitoring
Regular distribution audits on gender, university, and location in match results.
📜
PDPA Compliance
User consent required. Data retention policy. Audit trail for all access.
Scoring Model
Semantic Fit Score (SFS)
Explicit skill match 45%
Inferred skill relevance 25%
Learning Velocity 15%
Evidence strength 10%
Domain readiness 5%
Readiness Signal (RS)
Verified skills 30%
Learning Velocity 25%
Project evidence 20%
Leadership signals 15%
Career preference alignment 10%
Talentbank Career OS Integration

Lumina Graph is designed headless-first. Every engine exposes clean REST endpoints that plug directly into Talentbank's existing Career OS without requiring a platform rebuild.

GET /api/candidates/{id}/talent-profile
GET /api/candidates/{id}/readiness-score
POST /api/jobs/{id}/match
GET /api/universities/{id}/cohort-intelligence
POST /api/ai/extract-resume