🎯 Babigon's Competitive Strategy

The builder's blueprint. Technical architecture, demo plan, pitch structure, and execution roadmap for winning the BDI Hackathon 2026.

  Commander's Assessment: Best-in-class strategy document (1,680 lines). 2 rounds of competitive debate with Ginnie. All critical gaps now have executable plans. Score trajectory: 5.5 → 7.5 → 8.8/10.

1. Technical Architecture

3-Layer Data Engine → XGBoost Fusion → Risk Index

Layer	Data Source	Processing	Output
Layer 1	Sentinel-1 InSAR	LiCSBAS SBAS pipeline → Deformation velocity map	Subsidence rate (mm/yr), acceleration, time series stats
Layer 2	Dashcam footage	YOLOv8 → Damage classification + localization	Crack/pothole/depression scores per 100m segment
Layer 3	Environmental APIs	Rainfall, soil moisture, temperature, traffic load	Risk contribution from direct collapse triggers
XGBoost Fusion		24+ feature vector → Collapse Risk Index	0-100 risk score per 100m segment + SHAP explainability

  No existing system does this. Sinkhole Scanner (2021): InSAR only. InSARTrac (2023): CV+InSAR for structures only. RDD2022: dashcam CV only. SinkAlert is the first 3-modal fusion for road collapse.

2. ThaiLLM: The Patipan Strategy

4-Model Agentic Pipeline

Judge Patipan Prasertsom (ThaiLLM creator) will score ThaiLLM usage with zero tolerance for superficiality. Our response:

Model	Role	Why
1️⃣ ThaiLLM-8B-ToolUse	Route citizen LINE reports	99.9% tool routing accuracy. Fine-tuned on 5 road-safety tools.
2️⃣ Typhoon-S	Generate Thai alerts	32K context, hermes tool parser, sovereign training.
3️⃣ THaLLE-0.2	Complex reasoning	Best Thai benchmarks. Thinking mode for multi-factor analysis.
4️⃣ Playground API	Final synthesis	BDI's own infrastructure. Direct connection to their ecosystem.

ToolUse Fine-Tuning Plan (Critical Item #1)

5 road-safety tools: query_road_risk, generate_alert, classify_citizen_report, get_historical_incidents, recommend_inspection_schedule

200 training examples in JSONL format with Thai system prompts.

# Training command (Unsloth QLoRA, 45 min on A100)
python finetune_tooluse.py \
    --model_name "biodatadyne/ThaiLLM-8B-ToolUse" \
    --train_data "./data/tooluse_road_safety_train.jsonl" \
    --output_dir "./checkpoints/ThaiLLM-8B-ToolUse-RoadSafety" \
    --num_train_epochs 3 --use_peft true --lora_r 16

Deadline: June 29, 2026 | Owner: ML sub-team lead | Success: >90% tool accuracy on held-out test

3. Demo Day Plan

5-Minute Pitch Script (July 25 @ Lido Connect)

Time	Section	Content
0:00-0:30	Hook	"Bangkok sinks 10mm/year. Samsen Road: 50m sinkhole. ฿1,000M damages. We built the solution."
0:30-1:30	Architecture	3-layer system: InSAR ← Sentinel-1, Dashcam AI ← YOLOv8, Environmental ← ThaiLLM. XGBoost fusion.
1:30-2:30	ThaiLLM Demo	Live: "ตรวจสอบความเสี่ยงถนนแถวสุขุมวิท" → 4-model pipeline → Thai alert + SHAP chart.
2:30-3:30	LINE Bot	Scan QR code on screen → Citizen reports pothole → ThaiLLM classifies → Alert in DOH dashboard.
3:30-4:30	AWS + Couchbase	Bedrock Agent for autonomous monitoring. Couchbase vector search for similar-risk patterns.
4:30-5:00	The Ask	"7 THB/km. 72,556 km. First in ASEAN. Let us protect Thailand's highways."

"One Live Component" Strategy

Only the ThaiLLM alert generation runs live. Everything else is pre-computed with SHAP overlays. Backup: USB drive with static HTML + local video. If WiFi fails, we still deliver.

4. Competitive Differentiation

Competitor Profile	Odds	SinkAlert Advantage
PM2.5 Dashboard	40%	Crowded space. We offer 3-modal fusion they can't match.
Flood Prediction	25%	Single modality. Our cost disruption (7 THB vs 60K THB/km) is decisive.
Traffic Accident AI	20%	CV-only. No InSAR. No 4-model ThaiLLM pipeline.
Generic AI Safety Chatbot	15%	Basic ThaiLLM = losing special prize. Our fine-tuning + LINE Bot is unique.

  Judging Dimension Coverage: Each dimension maps to specific evidence in our pitch:

  Dr. Teeranee → data-driven national impact ✅

  Patipan → deep ThaiLLM integration ✅

  Benj → research methodology ✅

  Adisak → Bangkok urban data ✅

  Koravit → user-centered design ✅

  Sasawat → AWS Bedrock usage ✅

  Cheryl Lee → gov tech standards ✅

5. Execution Timeline

Week	Dates	Milestones
Week 1	Jun 16-22	✅ Verify InSAR coverage · Write 200 ToolUse examples · Begin dashcam collection
Week 2	Jun 23-29	✅ Fine-tune ToolUse · Process Bangkok deformation · YOLOv8 on Thai roads · AWS Workshop
Week 3	Jun 30 - Jul 6	Build risk dashboard · Interview road engineers · Compile incident database
Week 4	Jul 7-13	⭐ SOFT PITCH (Jul 10) · Integrate feedback · Begin LINE Bot
Week 5	Jul 14-20	LINE Bot complete · Couchbase vector search · Bedrock Agent · Pitch rehearsal
Week 6	Jul 21-25	⭐ DEMO DAY (Jul 25) · Final rehearsal · Judge Q&A prep · Deploy

6. Cost Model (7 THB/km)

Component	Annual Cost	Per km (72,556 km)
InSAR Processing (LiCSBAS pipeline)	฿30,600	฿0.42
Environmental APIs (Open-Meteo, TMD)	฿0 (free tier)	฿0.00
Dashcam Hardware (200 units × ฿2,000 ÷ 3yr)	฿133,300	฿1.84
XGBoost Inference (AWS SageMaker)	฿36,000	฿0.50
ThaiLLM API (BDI hackathon access)	฿0	฿0.00
Personnel (0.5 FTE engineer)	฿180,000	฿2.48
Cloud Infrastructure (AWS)	฿96,000	฿1.32
TOTAL	฿475,900	฿6.56 ≈ ฿7

  8,500× cheaper than GPR Raptor-45 (฿60,000/km). Even worst-case (฿15/km) = still 4,000× cheaper. The cost advantage is structural.