🕳️ SinkAlert

Bangkok Urban Risk Intelligence — Soft Pitch Jul 10, 2026

01 — The Problem

Bangkok sinks. Roads collapse. No warning system exists.

40+
Urban sinkholes since 2017
78%
Occur in rainy season (May-Oct)
3 days
DMR response time (manual)
2-3 cm
Annual subsidence rate

Current process: Citizen reports → DMR dispatches inspector (3 days) → manual assessment → paper report.
With SinkAlert: Dashcam → AI detection → instant alert (2 seconds).

02 — Tech Stack (All AWS)

Data Pipeline

Kinesis Video1 active stream
S3 Data Lake15,000+ objects
Sentinel-1 InSAR10 years, 12-day cadence
TMD Rainfall2018-2026

AI Models

YOLOv8 CV4-class crack detection
XGBoost Risk13 features
Bedrock NovaChat + Vision
SageMakerScript ready ▼ GPU pending
S3 Kinesis Bedrock SageMaker IAM YOLOv8 XGBoost Nova Pro Sentinel-1 TMD DMR Japan COCO

03 — End-to-End Pipeline

1
Ingest: Dashcam footage → Kinesis Video Streams → S3 bucket
2
Detect: YOLOv8 identifies cracks (D00-D40) — 4 severity levels
3
Analyze: Nova Vision on Bedrock — describes crack + risk factors
4
Predict: XGBoost model — 13 features → risk score with confidence interval
5
Alert: LINE/SMS to DMR engineers — Thai language, data-cited

Dashcam → Kinesis → S3 → YOLOv8 → Nova Vision → XGBoost → Alert
Latency: <2 seconds | Thai language | Free + Open

04 — Honest Performance

No inflated numbers. Here's the truth.

YOLOv8 (CPU Baseline)

mAP@500.077
Dataset6,221 images
Epochs15 (CPU limited)

→ GPU target: mAP 0.75+ (50 epochs, batch=32, imgsz=640)

XGBoost (Limited Data)

Bangkok sites10 (need 400+)
Current AUCdegenerate*
Validation5-fold CV + bootstrap

→ With full DMR: AUC 0.93+ (matching published SOTA)

16,669
Labeled crack boxes
4
Crack classes (D00-D40)
13
Risk features
1,000
Bootstrap resamples

* AUC = 1.0 with only 1 labeled province — the model learns "Bangkok has subsidence" = perfect separation. Honest pipeline catches this.

05 — Live Demo

Working now at aws.cfoth.ai

💬
Chat — Thai QA
📷
Image Analysis
⚠️
Risk Calculator
☁️
AWS Status

aws.cfoth.ai/pages/demo.html

06 — BDI × AWS Roadmap

What we need from BDI/AWS

✅ Bedrock NovaWorking
✅ S3 + KinesisWorking
🔴 SageMaker GPUQuota = 0
🔴 GLDAS GroundwaterNeed Earthdata

Week 1 after GPU grant

Day 1GPU train YOLOv8 (50 epochs)
Day 3Deploy SageMaker endpoint
Day 55 dashcams on Rama IV
Day 7LINE alert to DMR