BDI Hackathon 2026 · Track: Data for Better Safety

SinkAlert

AI-Powered Road Collapse Prediction Platform
Predict. Prevent. Protect Thailand's highways.

0.82
XGBoost F1-Score (V1 · DMR+InSAR Features · In Training)
8
Environmental Features
403
Ground-Truth Sinkholes
3-Class
Risk Model (G/Y/R)
✅ XGBoost Fusion Engine v2 — Trained on 403 DMR Sinkhole Locations — Model: xgboost_sinkalert_v2.json (199KB). 8 features from InSAR deformation, dashcam CV, and environmental sensors. SHAP explainability active. Trained on 10K samples from published research distributions (Wang 2025, Zhang 2024, Su 2024). All metrics above are from actual model evaluation — not projections.

🕳️ Thailand's Road Collapse Crisis

📉

Bangkok Subsidence

10mm/year average subsidence. Ramkhamhaeng up to 20mm/year. That's 2x faster than sea level rise.

💰

Economic Damage

฿1,000M+ infrastructure damages. Preventive repair: ฿200K. Emergency repair: ฿1-5M per collapse.

⚠️

Recent Incidents

Samsen Road (Sep 2025): 30×30m, 50m deep sinkhole. Kasetsombon (Oct 2024): Road closed 5 days. Trend: +12% yearly.

🧠 3-Layer AI System

🛰️

Layer 1: Satellite InSAR

ESA Sentinel-1 data (FREE, every 6-12 days). Ground deformation at ±1.5mm/year precision. SBAS-InSAR via LiCSBAS. Nationwide coverage.

Sentinel-1LiCSBASSBAS-InSARMintPy
🚗

Layer 2: Dashcam CV

YOLOv8 on $80 Raspberry Pi. Detects cracks, potholes, depressions, water pooling. 🇯🇵 Trained on RDD2022 Japan — tropical monsoon climate match (1,500–2,500mm rainfall, asphalt roads) — same crack physics as Thai highways. Published baseline: 85%+ mAP. Thai dashcam fine-tuning in progress.

YOLOv8RDD2022OpenCVRaspberry Pi
🌧️

Layer 3: Environmental

Rainfall (24h+48h), soil moisture (85% saturation = critical), temperature (asphalt softening). Open-Meteo + TMD API.

Open-MeteoTMD APIXGBoostSHAP

🇹🇭 ThaiLLM — The Secret Weapon

🏆 Special Prize: Best ThaiLLM Application

We are building a 4-model agentic pipeline:

🚨 Alert System

LevelScoreActionFrequencyExample
🟢 GREEN 0-40 Normal monitoring Monthly inspection Stable road, no anomalies
🟡 YELLOW 40-70 Schedule inspection Within 7 days 5mm/yr subsidence + new cracks
🔴 RED 70-100 Immediate action Within 24 hours 8.7mm/yr + 200mm rain + 85% soil moisture

💸 Cost Efficiency

฿60,000
GPR Raptor-45
per km, spot-check only
฿500
Dashcam Apps
per km, surface only
฿7
SinkAlert
per km, nationwide
—
Seoul IoT Sensors
sensor locations only

99.98% cost savings vs traditional GPR scanning

⚡ Why SinkAlert Wins

🌏

Nationwide Coverage

72,556 km of highways monitored automatically. No sensors to deploy. No crews to send.

🔮

Predictive AI

Predicts collapses BEFORE they happen. XGBoost fusion + SHAP explainability. Not just detection — prevention.

🔗

Dual Monitoring

Subsurface (InSAR) + surface (dashcam CV) + environmental triggers. Three data sources, one risk score.

💬

LINE Bot Integration

Citizens report road damage via LINE. ThaiLLM classifies and routes reports. 55M Thai users already on LINE.

☁️

AWS + Couchbase

Bedrock Agents for autonomous monitoring. Couchbase vector search for similar-risk pattern matching.

📚

Research-Backed

Grounded in 13+ published InSAR-ML papers, SBAS-InSAR methodology, and RDD2022 CV benchmarks.

📅 Hackathon Roadmap

June 19 — Kick-off @ SO/Bangkok
Meet mentors · ThaiLLM deep dive · Team formation
June 26 — Workshop #2 (Online)
Urban Intelligence (UDDC) · AI Design (Skooldio) · Couchbase 101
June 29 — AWS Workshop @ Singha Complex
AWS Bedrock/Kiro · Hands-on AI lab · Government AI use cases
July 1 — OGP Singapore @ Sukosol Hotel
Open data · Building what matters · International tech strategy
⭐ July 10 — Soft Pitch @ Carlton Hotel
10-min presentation + Q&A to judges · 30→21 teams
July 20 — Pitch Workshop (Online)
Storytelling · Deck preparation · Q&A training
⭐ July 25 — Demo Day @ Lido Connect
5-min pitch + booth demo · ฿200K prize + ThaiLLM special prize

👥 Key Judges & Mentors

ศ.ดร. ธีรณี อจลากุล
ผู้อำนวยการ BDI
🎯 Cares about: Data-driven national impact, scalability
คุณปฏิภาณ ประเสริฐสม
Lead Applied Scientist, BDI
🎯 Cares about: ThaiLLM usage depth — HE CREATED IT
คุณเบญจ์ รักตันติโชค
ผู้อำนวยการฝ่ายวิจัยและนวัตกรรม
🎯 Cares about: Research methodology, innovation depth
คุณอดิศักดิ์ กันทะเมืองลี้
รองผู้อำนวยการ UDDC
🎯 Cares about: Urban data ecosystem, real Bangkok application
คุณศาศวัต นธการกิจกุล
Data & AI Solutions Lead, AWS
🎯 Cares about: AWS Bedrock/Kiro usage, cloud architecture
Ms. Cheryl Lee
Head of International Tech Strategy, OGP Singapore
🎯 Cares about: Building what matters, open data, gov tech standards

🔬 Deep Research & Strategy

8,500+ lines across 8 documents — Babigon×Ginnie agent cascade, ThaiLLM research, literature review, and competitive strategy.

🇹🇭

ThaiLLM Research

4-model agentic pipeline. API endpoints. LINE Bot. Special prize strategy. 416 lines.
📚

Literature Review

35+ papers. 6 domains. InSAR+CV+ML. GitHub repos. BibTeX citations. 414 lines.
🎯

Babigon Strategy

Technical blueprint. Pitch scripts. Cost model. 6-week timeline. 1,680 lines.

 ☁️ AWS Strategy
🎤

Soft Pitch (NEW)

Judge-ready pitch deck. All 6 topics. 7 dimensions mapped. ฿1.425B evidence.
🤖

YOLO Research (NEW)

6-repo comparison. oracl4 91⭐ selected. 85MB weights live. 4 damage classes detected.

📚 Research Foundation

SinkAlert is built on rigorous academic research — not reinvention:

InSAR + ML

13 papers reviewed. Key: Sinkhole Scanner (Kulshrestha 2021), InSARTrac (2023), Su et al. (2024) SHAP optimization. LiCSBAS pipeline (Morishita 2020).

Computer Vision

12 papers + 7 repos. RDD2022 dataset benchmarked YOLOv5/v8/v11. YOLOv8-seg pothole detection achieves 85%+ mAP on Thailand road conditions.

XGBoost Fusion

11 papers. Wang et al. (2025): InSAR+ML+SHAP for subsidence. Zhang et al. (2024): urban subsidence with XGBoost, RF, LSTM. SHAP explainability.