Selisih is an analytics platform powered by machine learning that learns from thousands of historical BPK audit findings. Every month, it ranks the most suspicious APBD realisasi from 552 regional governments — so your DJPK team focuses on what truly needs investigation, not random sampling.
Synthetic example — anonymised SIKD data, no personal data stored.
Trillions of rupiah in potential BPK findings every year — not because your team isn't capable.
DJPK receives APBD realisasi reports from 552 regional governments every month. Manual checks and random sampling leave the more sophisticated patterns invisible. Anomalies are scattered across thousands of BAS account lines, and analyst teams are exhausted chasing samples that often aren't the most problematic.
The result: preventable BPK findings that slip through, delayed remediation recommendations, and an analyst workload that isn't human-scale. Selisih was built to fix all three — in one flow that feels familiar to DJPK analysts, not data scientists.
“With Selisih, I can focus on the Pemda that really need investigation — not random sampling every month.”
A machine-learning model picks the most urgent APBD realisasi every month — trained on historical BPK audit findings. Your team starts from the highest-value Pemda to investigate.
Behind every realisasi: three concrete reasons from AI analysis — budget absorption patterns, recording lag, Pemda fiscal tier. Ready to paste into an official memo.
SIKD rules live as a rule-based layer running alongside the AI model, referencing Permendagri 90/2019. Your team can add or adjust rules without a programmer.
Instead of starting anywhere, your team opens the most suspicious APBD realisasi for the month — already ranked by risk score, already labelled by BAS account.
Three concrete reasons per realisasi — budget absorption vs pagu, recording lag, Pemda fiscal tier — in sentences you can cite directly in an official memo.
| Pola yang diperiksa | Temuan |
|---|---|
| Serapan vs pagu anggaran | 122,65% (rata-rata 84%) |
| Lag pencatatan realisasi | 75 hari — persentil ke-91 |
| Tier fiskal Pemda | Tier B — menengah |
| Frekuensi koreksi LRA | — |
| Dokumen pendukung | Tidak dilampirkan |
Confirm as suspicious, reject, or escalate to the BPK team — with a short note. Every decision strengthens the model for next month.
A full official memo with absorption evidence and cited SIKD rules — the analyst just reviews and signs.
Keputusan auditor ditulis ke tabel anotasi · model retrain mingguan menggunakan label ini · versi model tersimpan untuk audit trail 5 tahun.
A workflow that used to take the whole week now fits into a few hours — with a complete audit trail behind every decision.
At the core of Selisih: classical machine learning with decades of research behind it — deterministic, auditable, runs without internet. On top: five optional AI-copilot features (generative AI for narratives, RAG for regulations, NLP for queries) that help your team move faster. The copilot can be switched off at any time; the core keeps running. The final decision always belongs to the DJPK analyst.
'Realisasi from Pemda X with high absorption anomalies in Q3' — the queue filters itself. No query language to learn.
Three technical reasons condensed into one paragraph — ready to send to the Pemda or supervisor as a finding note.
Compares BAS account codes in realisasi against the budget classification — helps analysts catch potential misstatements that routinely go unnoticed.
'What are the conditions for realisasi exceeding pagu?' — a brief answer with citations from Permendagri 90/2019 and UU 1/2022 HKPD.
The system prepares the memo; the analyst approves. All absorption evidence and SIKD rule citations are already attached.
↳ AI only helps explain and compose. Scoring and decisions stay with your team.