Lompat ke konten
Jakarta, Indonesia, UTC+7 Saat ini CTO di Sxored
EN ID
ALM ALM Abdul Latif Munjiat
ALM ALM
EN ID
Kembali ke studi kasus
01 / Sxored

Document intelligence banks can audit.

I run the credit intelligence platform that turns bank statements, credit history, KYB filings and property appraisals into structured data lenders can act on. Event driven Go microservices with RabbitMQ and Temporal, a React and TypeScript SaaS plus back office, on self managed Kubernetes across AWS and GCP.

Role Chief Technology Officer
Period Sep 2024 to Present
Customers Banks, lenders, corporates
Certification ISO 27001
01 / Masalah

Documents that decide credit

A lender's decision rests on documents that were never designed to be read by machines: scanned statements from dozens of banks, credit history in fixed width layouts, appraisals with tables that break across pages.

Manual review is slow and inconsistent. Naive automation is worse: an extraction that is confidently wrong sends a bad file to a credit committee, and nobody can point to where the number came from. The requirement was never raw accuracy. It was accuracy with provenance, under supervision, at a cost per document that a lender will keep paying.

02 / Yang saya pegang

Platform, team, and the compliance posture

Architecture Durable extraction workflows, queue isolation per document class, and a review path for anything below confidence.
Team Hiring, on call, and the review culture. Small, senior, high autonomy, no heroics rewarded.
Cost Cost per document as a first class metric, self managed Kubernetes where it pays for itself.
Trust Data handling, retention and audit trails that survive a bank security review.
03 / Keputusan dan dampak

Keputusan yang saya pertahankan

Each one traded something away. That is the point of writing them down.

KeputusanDampak bisnis
Durable workflows per document class instead of one generic pipeline A failing appraisal parse cannot stall statement processing, and each class can be tuned and priced on its own.
Confidence thresholds with a human review path, rather than full automation Reviewers see only the uncertain fields, and every published value keeps a trace back to the page it came from.
Self managed Kubernetes over fully managed inference services Cost per document stays predictable as volume grows, at the price of owning the upgrade path ourselves.
Data residency and retention rules designed before the first bank security review Procurement conversations start from a document instead of a promise, which shortens every enterprise deal.
Quantified outcomes are pending customer approval. I will publish them here when they are cleared, not before.
Resilient systems over heroics: circuit breakers, backpressure and graceful degradation designed in, rather than patched in after incidents.
How I think about it
04 / What is next

Still shipping

The next year is about widening document coverage without widening the failure surface, and pushing more of the review loop back into the product so a reviewer only sees what actually needs a human.

Bicarakan karya ini Semua studi kasus