AEAP 1.0 Documentation · Advance Retro Solution Sdn Bhd

Welcome to the ARS Edge Analytics Platform

The ARS Edge Analytics Platform (AEAP) is a unified analytics environment for Pangkalan Data Utama. With AEAP you can prepare and govern data, explore it visually, build interactive reports, train and deploy machine-learning models without writing code, and automate high-volume eligibility decisions — all inside one governed platform with single sign-on and full audit.

AEAP is delivered as a subscription-licensed product, deployed on Kubernetes inside the PDSA environment. It has no external cloud dependency: every service, model, and dataset runs on infrastructure controlled by the Ministry.

What's new in AEAP 1.0

• NoCode ML leaderboard — five algorithm families trained in parallel with automatic champion selection on validation AUC, fairness checks across strata, and one-click deployment to the serving layer.
• Decision banding — configurable score thresholds route borderline cases to human review instead of auto-decline.
• Auto-insights — every exploration computes highest/lowest, mean, and spread narratives automatically.
• Licence metering dashboard — real-time decision-volume consumption against contract entitlement, exportable for audit.
• Bahasa Melayu UI pack — full BM localisation ships with 1.0; toggle per user under Profile → Language.

Get started

Sign in

Browse to the platform URL and sign in with your Ministry credentials. AEAP delegates authentication to Keycloak, which federates to the Ministry's Active Directory over SAML 2.0. Multi-factor authentication follows your directory policy. Your role (Admin, Steward, Analyst, or Viewer) determines which modules and datasets you see.

Your first ten minutes

1. Home shows platform-wide indicators and your recent items.
2. Open Data and click any dataset to see its column profile and quality score.
3. Click Explore this dataset, pick a measure, and switch chart forms — the chart redraws instantly.
4. Open Reports and view the National Socioeconomic Overview; every chart is live.
5. Open NoCode ML to see a finished training run, then follow the champion into the Decision Engine.

Data catalogue

Every dataset in AEAP is registered in the catalogue with an owner, source system, refresh schedule, quality score, and classification tags. Datasets are tiered gold (governed, SLA-backed), silver (curated), and bronze (raw landings, visible to Stewards only).

Click a dataset to open its detail page: row/column counts, per-column types, null rates, distinct counts, and assigned analytical roles (Key, Category, Measure, Date). The Null rate chart flags columns above the 3% imputation-review threshold.

Quality scores

The DQ engine recomputes dataset scores daily as a weighted blend of completeness, validity, consistency, freshness, and uniqueness. Scores below 92% raise a Governance alert and appear amber/red throughout the UI.

Preparing data

From a dataset page, Open in notebook launches a JupyterLab session with the dataset mounted read-only via DuckDB, or use the visual preparation canvas to chain steps — filter, derive, join, pivot, impute — that compile to SQL and run in-database. Preparation outputs are registered back into the catalogue with lineage to their sources.

# Example: the same preparation, expressed in the notebook
import duckdb
con = duckdb.connect("padu.duckdb")
b40 = con.sql("""
  SELECT negeri, COUNT(*) AS isi_rumah,
         MEDIAN(pendapatan_kasar) AS pendapatan_penengah
  FROM padu_isi_rumah
  WHERE kumpulan_pendapatan = 'B40'
  GROUP BY negeri ORDER BY 3
""").df()

Connectors

Exploring data

Explore & Visualize is the fastest path from question to chart. Choose a dataset, a dimension, and a measure; AEAP profiles the working set and draws a ranked bar, scatter, or share chart. Switching the measure or chart form redraws without a page reload.

The Auto-insights panel narrates the extremes, mean, and spread of the current view. In the full platform, the Explain panel extends this with decomposition and correlated-factor analysis, and Pin to report places the current chart on a report canvas.

Creating reports

Reports are multi-page canvases of linked objects. Edit mode offers a drag-and-drop palette; every object binds to a catalogue dataset and inherits its refresh schedule. Page-level filters and cross-object actions (click a state → every object filters to it) are configured without code.

• Refresh — report data refreshes with its underlying datasets; viewers always see the refresh timestamp.
• Scheduling — distribute any report as PDF or XLSX on a cron schedule to email or a shared drive.
• Row-level security — dataset RLS rules follow the viewer, so one report serves many audiences safely.

Report objects

Sharing & export

Share a report by link (respecting RLS), export any page to PDF, or export any table object to XLSX/CSV. Exports are watermarked with the viewer's identity and logged to the audit trail. Printing uses a light print stylesheet automatically.

NoCode ML

NoCode ML takes a governed dataset to a deployed model in five guided steps:

1. Data — pick the training dataset and an optional holdout policy (default: 30% stratified).
2. Target & features — choose the target column; AEAP proposes features and excludes leaky ones.
3. Auto-EDA — distributions, correlations, missingness, and target balance, generated automatically.
4. Model comparison — five algorithm families (logistic regression, decision tree, random forest, gradient boosting, neural network) train in parallel; the leaderboard ranks them by AUC with KS, F1, and Gini alongside. ROC curves and feature importance are one click away.
5. Deploy — register the champion to the model registry and expose a scoring endpoint on the serving layer with autoscaling and drift monitoring (PSI).

Decision Engine

The Decision Engine executes decision flows: directed graphs of rule sets, model scores, and banding nodes. The shipped STR eligibility flow runs hard exclusion rules first (fast, cheap, explainable), scores the remainder with the champion model, then bands the score into approve / review / decline.

• Throughput — engineered for 200 decisions/second sustained; the contract volume of 1M decisions/year is metered on the Governance page.
• Champion/challenger — a challenger model can shadow-score any percentage of live traffic; agreement is tracked before promotion.
• Audit — every decision stores its inputs, score, triggered rules, outcome, latency, and flow version for 7 years in WORM storage.
• Batch — the same flow re-scores changed households nightly; batch volume is metered separately and unlimited.

POST /v1/decisions/str-eligibility
{ "subject": "IR-88412-31", "context": { "pendapatan_kasar": 3120, "bil_tanggungan": 4 } }

→ 200 OK  { "outcome": "APPROVE", "score": 0.9412, "flow": "v4", "latency_ms": 22,
            "audit_id": "D-2026-084412" }

Pipelines

Pipelines move and transform data on schedules or streams. Author them visually in KNIME or the built-in designer; AEAP compiles them to Airflow DAGs for execution. Each pipeline carries an SLA; the lineage view traces any gold dataset back to its source extracts. Failures page the on-call Steward and appear on the Governance dashboard.

Integrations

AEAP integrates with the PADU estate at three tiers. The inventory is identical whether the analytics engine is AEAP or a commercial suite — the interfaces below do not change.

Tier 1 — Direct runtime integrations

Tier 2 — Platform & security dependencies

Tier 3 — Observability & assurance hooks

Enterprise BI — Microsoft Power BI

Five Enterprise BI seats are delivered as Microsoft Power BI alongside the platform. The trial's Enterprise BI page embeds an equivalent interactive report so the workflow can be exercised before licences are provisioned. The production integration:

• Connectivity — DirectQuery/ODBC to PADU PostgreSQL and the platform's semantic views; import mode with scheduled refresh for heavy aggregates. The on-premises data gateway runs on the dedicated Windows Server (VCF vSphere).
• Security — row-level security bound to the same AD groups as platform RBAC; PII stays masked at the database view layer so BI seats inherit PDPA controls automatically.
• Distribution — licensed seats author and publish; consumers view via portal embed or PDF/PPT export. Platform dashboards remain the operational layer; Power BI serves ad-hoc and executive analysis.

Administration

Deployment topology

Production runs on RKE2 Kubernetes — 3 control-plane and 4 worker nodes (16 vCPU / 64 GB each), 4 TB NVMe block and 10 TB object storage — behind the Ministry WAF. All services deploy from a Helm umbrella chart via GitOps (ArgoCD); helm install aeap brings up a working platform in under 30 minutes. Scaling out is adding worker nodes.

Roles

Backup & disaster recovery

Velero snapshots all cluster state and volumes nightly to off-cluster object storage. The documented restore runbook targets RTO 2 hours / RPO 24 hours; the last restore drill completed in 1h 41m. Runbooks ship in Bahasa Melayu and English.

Security & PDPA

• TLS 1.3 for all traffic; AES-256 at rest (ZFS native encryption and MinIO SSE).
• SSO via Keycloak → Ministry AD (SAML 2.0) with MFA; no local passwords in production.
• PII columns are tagged under PDPA classification and masked for the Analyst role by default; unmasking requires Steward approval and is audit-logged.
• CIS Kubernetes benchmark enforced in CI (kube-bench); images scanned with Trivy on every build.
• External penetration test on each major release; findings and remediation evidence available to evaluators.

REST API

Everything the UI does is available over REST with OAuth2 client-credentials. Interactive OpenAPI documentation is served at /api/docs on the platform host.

Accessibility

AEAP targets WCAG 2.1 AA: full keyboard navigation with visible focus states, ARIA labelling on charts and diagrams, colour palettes chosen to survive the common colour-vision deficiencies, and respect for the operating system's reduced-motion preference. Report authors get an accessibility checker that flags contrast and missing-label issues before publishing.

AEAP 1.0.0-trial documentation · © 2026 Advance Retro Solution Sdn Bhd · All product data on the trial server is synthetic. Document structure follows the conventions of leading visual-analytics suites for evaluator familiarity.