Migration Strategy & Architecture Plan
Leveraging Amazon Bedrock, IoT Core, SageMaker & Kiro for Maximum Hackathon Impact
| # | Dimension | Weight | Current | With AWS | What AWS Adds |
|---|---|---|---|---|---|
| 1 | Problem Fit & Use Case | 20 | 14 | 18 | AWS IoT architecture shows real-world deployment; KVS for dashcam integration |
| 2 | PoC Progress & Demo | 20 | 12 | 17 | Bedrock AgentCore demo = production-ready; SageMaker notebook = training proof |
| 3 | Data & AI Approach | 15 | 11 | 14 | Bedrock multi-model + SageMaker XGBoost + Rekognition CV = AWS-native AI stack |
| 4 | Validation & Evidence | 10 | 5 | 7 | Bedrock Guardrails for safety; CloudWatch for monitoring |
| 5 | Impact, ROI & Social Value | 15 | 12 | 14 | Serverless cost model (Lambda); AWS Free Tier for pilot; transparent pricing |
| 6 | Team & Learning | 10 | 6 | 8 | Kiro adoption shows AWS tool proficiency; spec-driven development discipline |
| 7 | Ecosystem & Strategic Fit | 10 | 6 | 9 | 100% AWS-native: Bedrock + SageMaker + IoT + Lambda + KVS = runs on sponsor's platform |
| TOTAL | 100 | 66 | 87 | +21 points from AWS alignment | |
| Model | Use Case | Why This Model | Cost |
|---|---|---|---|
| Claude 3.5 Sonnet | Thai-language risk reasoning, alert generation | Best Thai reasoning among Bedrock models. Extended access free for hackathon evaluation. | $6/M input · $30/M output tokens |
| Amazon Nova Pro | Multimodal: satellite image + text analysis | AWS-native. Processes InSAR visuals + Thai descriptions together. | AWS on-demand pricing |
| NVIDIA Nemotron Super 49B | Research synthesis, 1M context window for reading papers | Already using via NVIDIA NIM. Available on Bedrock Marketplace (p.18). | Marketplace pricing |
| DeepSeek R1 | Complex root cause analysis | Already using via DeepSeek API. Also on Bedrock Marketplace. | Marketplace pricing |
Why this wins points: AgentCore is AWS's flagship agent platform presented at BDI Workshop #4. Using it shows you listened to the sponsor's technical direction.
| Component | SinkAlert Use | Score Impact |
|---|---|---|
| AgentCore Runtime (p.22-23) | Multi-agent: Supervisor agent (Strands) orchestrates InSAR agent + CV agent + Alert agent | D2 (+3pts): Production-grade multi-agent architecture |
| AgentCore Memory (p.22) | Persist risk assessments, incident history, Thai-language context across sessions | D3 (+1pt): Shows memory-aware AI design |
| AgentCore Identity (p.24-26) | Secure access to AWS resources (S3, SageMaker) and external tools via OAuth | D4 (+1pt): Secure, auditable agent operations |
| AgentCore Observability (p.27-28) | CloudWatch traces, OTEL logs. Full audit trail of every AI decision. | D2 (+2pts): Operational readiness visible to judges |
| AgentCore Gateway (p.21) | Browser tool for automated data collection. Code interpreter for analysis. | D3 (+1pt): Agent tooling maturity |
| Feature | SinkAlert Use | Proven Metric (from AWS) |
|---|---|---|
| Knowledge Bases | RAG over Thai disaster documents: DMR sinkhole maps, DOH road specs, DDPM protocols | Contextual retrieval with citation |
| Content Filtering | Block harmful/misleading alerts before they reach the public | 88% harmful content filtered |
| Automated Reasoning | Verify model outputs against ground truth — prevent hallucinated sinkhole alerts | 99% hallucination detection accuracy |
| PII Redaction | Strip personal data from incident reports before public dashboards | HIPAA-eligible PII masking |
| Groundedness Checks | Ensure every alert cites specific data sources (InSAR image, rainfall reading, road segment) | Context-based verification |
| Capability | Current Setup | AWS Migration |
|---|---|---|
| XGBoost Training | Local Python script (not trained yet) | SageMaker XGBoost built-in algorithm → managed training job with DMR labels |
| InSAR Processing | Local MintPy + ISCE2 | SageMaker Processing job → containerized ISCE2 pipeline → outputs to S3 |
| Model Registry | None | SageMaker Model Registry → versioned models with lineage tracking |
| Inference | Not deployed | SageMaker real-time endpoint → REST API for risk scoring |
| Hyperpod (p.5) | N/A | Future: distributed training on Trainium for large geospatial models |
RDD2022 provides multi-national road damage images. We selected Japan COCO as our transfer-learning source because it's the only subset with Thailand-matching road degradation physics:
| Condition | Japan | Thailand | Match |
|---|---|---|---|
| Climate | Tropical monsoon | Tropical monsoon | ✅ |
| Rainfall | 1,500–2,500mm | 1,200–2,400mm | ✅ |
| Road surface | Asphalt | Asphalt AC 60/70 | ✅ |
| Damage types | Alligator, pothole, crack | Alligator, pothole, crack | ✅ |
Norway (subarctic ❌), Turkey (dry ❌), USA (concrete roads ❌) were eliminated. Training on Japan's tropical monsoon crack patterns ensures transfer learning preserves relevant features before fine-tuning on Thai dashcam data.
Why this differentiates: The AWS AI Portfolio dedicates 5 full pages to IoT + KVS. This is a core AWS competency. SinkAlert positioning as an "IoT-connected road monitoring system" — even with simulated sensors — shows architectural vision.
| Component | SinkAlert Use | Demo Potential |
|---|---|---|
| Kinesis Video Streams | Dashcam/security camera ingestion → ML inference pipeline (p.10 diagram) | Show a video feed → KVS → Rekognition detecting road anomalies live |
| AWS IoT Core | Simulated soil moisture sensors → MQTT telemetry → IoT Events trigger | Dashboard showing "live" sensor data feeding the risk model |
| IoT Greengrass | Edge CV on roadside cameras — YOLOv8 runs locally, only alerts sent to cloud | Architecture diagram showing edge-to-cloud pipeline |
| IoT Device Shadow | Road segment "digital twin" with current risk state | Map click → device shadow shows full risk profile |
| Service | Monthly Cost (Pilot) | What It Does |
|---|---|---|
| Lambda | $0 (1M free requests/mo) | Alert processing, data transformation, API endpoints |
| S3 | ~$3/mo (50GB InSAR data) | Data lake: satellite imagery, sensor telemetry, incident reports |
| Bedrock (on-demand) | ~$15/mo (est. 500K tokens) | Claude 3.5 Sonnet for risk reasoning + Nova for multimodal |
| SageMaker (training) | ~$5/job (ml.m5.xlarge × 2hr) | XGBoost training on DMR labels |
| Kinesis Video Streams | ~$2/mo (1 stream, 1 cam) | Video ingestion for PoC demo |
| TOTAL Monthly (Pilot Phase) | ~$25/month | |
Kiro was featured at Workshop #3 and teams were explicitly instructed to download Kiro before the session. This is not optional — it's an expected tool.
| Kiro Feature | What It Does | SinkAlert Use |
|---|---|---|
| Spec-Driven Development (p.52-53) | Turn prompts into requirements, system design, discrete tasks | Generate SinkAlert architecture spec → Kiro agents implement scaffolding |
| Agent Hooks (p.54) | Auto-trigger agents on file save — generate docs, tests, optimize code | Auto-generate documentation when model code changes |
| Steering Files (p.55-60) | Markdown rules in .kiro/steering/ — version-controlled team knowledge | Code SinkAlert coding standards, AWS conventions, BDI hackathon rules |
| MCP Servers (p.61-65) | Connect AI to Jira, GitHub, AWS, databases | Connect Kiro agents directly to AWS Bedrock + SageMaker |
| Agent Skills (p.70-73) | Packaged task workflows (SKILL.md + scripts + references) | Create "SinkAlert Model Training" skill — reproducible ML pipeline |
| Custom Agents (p.74-77) | Specialized AI personas with scoped tools | Development agent (full access) + Production agent (read-only alerts) |
| Risk | Mitigation |
|---|---|
| AWS credits not activated in time | Use AWS Free Tier ($0) for all PoC services. Bedrock on-demand models incur cost only when called. Lambda free tier: 1M requests/month. |
| Bedrock model access delay | Request Claude 3.5 Sonnet Extended Access immediately (12-24hr approval). Fallback: use Amazon Nova (instant access, no approval needed). |
| SageMaker training results poor | Pitch the pipeline, not the accuracy. "XGBoost V1 trained on DMR labels — score improving weekly as we add InSAR features." Honesty > fake numbers. |
| AgentCore too complex for 8 days | Demo the simplest agent flow: User asks → AgentCore calls SageMaker endpoint → Bedrock explains result in Thai. 30-min setup. |
| # | Topic (5 min total) | AWS Mention | Time |
|---|---|---|---|
| 1 | ปัญหา & กลุ่มเป้าหมาย | "AWS IoT + KVS ingest dashcam data from 68,000km — the same architecture on p.10 of the AWS AI Portfolio" | 40s |
| 2 | ทางแก้ & Value Proposition | "Running on Amazon Bedrock AgentCore — cost ฿7/km, $25/mo pilot. 1,386 years of operation = cost of preventing one Samsen" | 50s |
| 3 | Data & AI Approach | "SageMaker XGBoost on DMR labels. Bedrock Claude 3.5 for Thai reasoning. Bedrock Guardrails: 99% hallucination detection. Rekognition CV augmentation." | 60s |
| 4 | PoC Progress | "Training running on SageMaker. AgentCore demo live at dashboard.cfoth.ai/sinkalert/. Built with Kiro AI-powered IDE — spec-driven development." | 50s |
| 5 | Feasibility & Validation | "AWS Free Tier covers entire pilot. S3 stores all Thai government data. Lambda handles 1M alert requests/month at $0." | 40s |
| 6 | ทีม & Next steps | "Post-Soft Pitch: migrate to Bedrock AgentCore production. Enroll in AWS Mentoring Hour. List on AWS Marketplace for Thai government procurement." | 40s |