Site Intelligence
We built this because spreadsheets and WhatsApp groups don't scale past 3 concurrent projects. This is what happens when you give construction data to prediction models trained specifically on Indonesian project patterns.
Currently in beta with 12 active sites across Jabodetabek.
What it does
Six modules, each solving a specific problem we kept running into on our own projects.
Delay prediction
Trained on 3 years of our own project history plus public procurement data from LPSE. Flags risk of schedule slip 2-3 weeks ahead. Not perfect — currently ~82% precision on our test set — but enough to trigger early mitigation.
PPE detection
Runs YOLO v8 on existing site CCTV — no new cameras needed. Checks helmets, vests, boots. We get ~15fps on a single RTX 3060 handling 4 feeds. Sends WhatsApp alert to safety officer when violation detected.
Environmental sensors
ESP32-based nodes measuring temperature, humidity, wind, noise. Auto-triggers work stoppage notification when wind > 40km/h or heat index > 35°C. Costs ~Rp 800k per node — we build them ourselves.
Cost forecasting
Tracks material prices from our supplier network (currently 23 suppliers in Jabodetabek). Predicts cost changes using seasonal patterns — useful for locking in concrete and rebar prices before spikes.
Equipment GPS
OBD-II dongles on excavators and dump trucks. Shows location, idle time, fuel consumption. The idle-time alerts alone saved one client Rp 47M/month in wasted rental fees.
Progress tracking
Weekly drone flights compared against BIM model. Calculates % completion per zone automatically. Still requires manual QC — the model sometimes miscounts rebar in dense areas — but cuts reporting time from 2 days to 3 hours.
How it's built
We run this on a pretty standard modern stack. Nothing exotic — reliability matters more than novelty.
┌───────────────────────────────────────────────────────────────┐ │ CLIENTS │ │ Next.js SSR │ React SPA │ PWA (field workers) │ REST API │ ├───────────────────────────────────────────────────────────────┤ │ API GATEWAY · rate limit · JWT verify · RBAC │ ├────────────┬────────────┬────────────┬────────────────────────┤ │ Project │ Material │ Equipment │ ML Inference │ │ Service │ Service │ Service │ Service │ │ (Node/TS) │ (Node/TS) │ (Node/TS) │ (Python/FastAPI) │ ├────────────┴────────────┴────────────┴────────────────────────┤ │ Redis Pub/Sub · event bus · cache invalidation │ ├────────────┬────────────┬─────────────────────────────────────┤ │ PostgreSQL │ InfluxDB │ S3 (files, drone images, ML models) │ │ (primary) │ (sensors) │ │ ├────────────┴────────────┴─────────────────────────────────────┤ │ INFRA: ECS Fargate │ ALB │ CloudFront │ Mosquitto MQTT │ └───────────────────────────────────────────────────────────────┘
Current system metrics
~8,000
events/sec peak
340ms
p95 inference latency
99.4%
uptime (last 90 days)
2.1 TB
time-series data stored
Security posture
Encryption
AES-256 at rest, TLS 1.3 in transit
RBAC
Per-project, per-role permissions
Audit log
Every action logged, 3-year retention
Backups
Daily automated, cross-AZ replication
Data residency
ap-southeast-1 (Singapore)
Compliance
Working toward SOC 2 (target Q1 2027)
Known limitations (we're working on these)
- • Delay prediction accuracy drops below 70% for projects under 3 months duration
- • PPE detection struggles in low-light conditions (before 6am, after 6pm)
- • GPS tracking requires cellular signal — doesn't work well in remote mountain sites
- • Cost forecasting currently limited to Jabodetabek supplier data
Want to try it on your site?
Site Intelligence is currently available as a beta add-on for BangunHub Pro and Enterprise. We onboard 2-3 new sites per month.
PT Karya Bangun Semesta · Engineering Division · Jakarta