Systems in production, and how they were built.
Three case studies in depth, then the rest of the deliveries. Client names appear only where they are cleared.
Plate recognition for Korlantas Polri
Part of Percik, the platform for vehicle registration and compliance checks run by the Indonesian National Police Traffic Corps. Plate recognition and vehicle identification across 5 sites nationwide.
- Accuracy
- 98.7% plate reading in live operation, measured on a human-checked sample of live reads
- Deployment
- Fully on-premise, on Proxmox/KVM virtualization I operated. No cloud.
- Models
- I trained the plate models and the detector and reader for metal-stamped VIN and engine numbers, including tamper detection for re-stamped or ground-off numbers.
- Field team
- 21 people (staff and contractors) in 7 teams of 3, SOP-driven daily reports, weekly coordination
- Validation
- Held-out test sets and client acceptance tests before go-live; live accuracy tracked and models retrained on site data afterwards
A plate reader that scores well on a test set can still fail in the field. The reads that mattered were the live ones, so accuracy was measured on human-checked samples of real traffic at the sites, and the models were retrained on what the sites actually saw.
Stamped VIN and engine numbers are harder than plates: low contrast, rust, paint and deliberate tampering. The reader flags numbers that look re-stamped or ground off.



Faces and uniform insignia are blurred.
Access control and daily health screening at an Indonesian refinery



The background, signage, screens and people are blurred.
Delivered by PT Adhya Arutala Amerta jointly with ARSA Technology, as a subcontractor under a main contractor. I led the 33-person team end to end: needs discovery, solution design, implementation, deployment and after-sales.
It screens about 10,000 employees and contractors. The final gate opens only when every check has passed:
- Five bare turnstiles, retrofitted with a computer, screen, camera and software: a card tap plus face verification.
- Ten custom-built self-service health kiosks: a camera-based Romberg balance test, blood pressure, heart rate, oxygen saturation and body temperature, each result tagged to the person by face.
- A 16-zone walk-through metal detector, extended with a computer, face recognition and a screen showing where a metal item was found.
- The final turnstile: a card tap.
The site has no internet access, so the whole system runs offline. Two-way integration with Honeywell Pro-Watch, the site's access control and directory, was worked out in-house with no vendor support and no API documentation.
AI agents in production, on-premise
A self-hosted stack I built and run in my company's own operations: LLMs served with Ollama and vLLM, and RAG over SOPs, the product and price-list knowledge base, active projects and the full CRM pipeline.
Workflow automation drafts quotations and full solution proposals (timeline, UAT checklist, statement of work), runs price-margin-cost analysis, sends follow-up reminders, answers staff questions, summarises client messages and enforces process steps.
An MCP server gives AI agents the CRM, quotations and proposals, the product catalog, projects, meetings, SOPs and client message threads, with a human-approval step before anything reaches a customer, a calendar or an invoice.
| Task | before | after |
|---|---|---|
| Quotation | 5 h | 30 min |
| Solution proposal (timeline, UAT checklist, SoW) | 25 man-h | 3 man-h |
| Price-margin-cost analysis | 7 h | 2 h |
| Process enforcement | 5 man-h | 30 min |
More deliveries
Restricted-area protection, Ministry of Defense
Facial recognition and AI video analytics ("Arsaca Eagle Eye") for perimeter security and access control at defense facilities, for Pushansiber Kemhan RI. Classified-handling procedures applied.
ARSA Face SDK
I hand-coded the original Python codebase, then set the new architecture and rewrote it in Rust. 99.45% face verification accuracy on LFW (10-fold, end to end); the edge model scores 99.18% on LFW at 13 MB, with no GPU.
Offline ML data platform
An air-gapped in-house platform for labelling, dataset versioning and training, with a self-hosted vision-language-model pre-labelling tool. In use for about two years.
Industrial safety analytics
PPE detection, activity recognition and heavy-equipment monitoring for multiple state-owned enterprises. The pictures below are ARSA's Basic Safety Guard product screens (demo data) and a frame from ARSA's demo footage.




ARSA AI Box series
An edge-AI platform that turns existing CCTV into real-time detection, fully on-premise. Productized for HSE compliance, traffic analytics, retail analytics and audience measurement.


Video analytics proof of concept, BRI (2023)
A multi-function video analytics proof of concept for Bank Rakyat Indonesia at the Pondok Indah branch, Jakarta.


Hardware and embedded




- About 150 RFID productivity terminals for PT Ungaran Sari Garments / Busana Apparel GroupFirmware, hardware manufacturing, backend2020
- Embedded voice-to-text engine for BotikaBuilt on Mozilla DeepSpeech2019
- Automated industrial gate with long-range UHF RFIDFor a logistics operator in East Java
- Radiology image analysis R&D at PT Adhya Arutala AmertaChest X-ray, CT/MRI; research and prototype stage only
- Commercial Face Recognition APIPublished on RapidAPI