Muhammad Ali Akbar Al-Qahri

Problem. Provincial poverty data spans changing administrative coverage and measurement windows, so direct rankings can hide data-universe differences and geographic clustering.

Approach. Harmonized BPS records into 38/34/32-province analytical universes, validated a 50% naive lag-1 and 50% Ridge ensemble with walk-forward testing, measured spatial autocorrelation, and packaged the results in a Next.js/Vinext portal.

Impact. National trends, province profiles, four spatial views, model diagnostics, and experimental 2026 forecasts are now inspectable in one interface without presenting projections as official BPS figures.

Problem. National SPPG records are difficult to compare when geographic coverage, address quality, and location precision vary across regions.

Approach. Built a responsive Leaflet portal with marker clustering, shareable filters, readiness scoring, province comparison, demand-gap methods, and an optional Supabase admin layer protected by RLS and audit logs.

Impact. Distribution, coverage concentration, data issues, and operational context can be inspected in one public interface while preserving a static fallback for reliable access.

Problem. Answer-key retrieval and prepublish QC were split across browser scripts and manual files, creating stale-key and schema-mismatch risks.

Approach. Unified two browser agents in one Playwright worker, passed keys through SQLite instead of intermediate files, added per-package checkpoints, resume controls, local monitoring, and reproducible reporting.

Impact. 33,500+ questions verified before publish across 600+ package runs, with 64K+ answer keys retrieved automatically instead of passed around by hand in CSV files.

Problem. Writing fresh practice packages by hand for many exam formats was slow, repeated items crept in, and every hand-typed key was a chance for an error to reach learners.

Approach. Built deterministic generators where the same seed reproduces the same package, with validators that re-derive keys instead of trusting the generator, added anti-duplicate checks against reference banks, and used Claude API plus a rule engine to draft TKA explanations behind a blind-review UI.

Impact. Validated at scale: 4,500 attention-test items across 50 packages with zero repeats, 3,000 diagrammatic items with zero repeated images, and 2,422/2,422 TKA keys independently re-verified.

Problem. Regional performance reporting was assembled by hand from multiple source systems — slow, inconsistent, and hard to audit.

Approach. Designed multi-database ETL into a star-schema warehouse and rebuilt reporting as four production Power BI dashboards, documented across 120 internship tasks.

Impact. Reporting time fell about 40% across five regional branches, and eight strategic dashboards reached 85% stakeholder satisfaction with direct input to C-level decisions.

Problem. Indonesia's higher-education records are scattered and regionally inconsistent, making institutions hard to compare across provinces and cities.

Approach. Built a Node.js pipeline that maps 5,433 institutions to official Kemendagri boundaries, guarded by 34 independent data-quality checks, shipped as an offline-ready Leaflet dashboard.

Impact. Accreditation, cost medians, and coverage for 38 provinces and 514 cities/regencies are explorable in one map — every figure traceable to a validated pipeline.

Problem. Provincial food-security planning needs forward-looking risk signals, but the underlying indicators are noisy, incomplete, and rarely validated.

Approach. Trained a Random Forest forecaster under time-series cross-validation with automated data validation, then exposed forecasts, feature importance, and provincial risk scores in Streamlit.

Impact. Reached R² 85.1% across 34 provinces — planners get an inspectable early-warning view instead of a static spreadsheet.