🌙 SinkAlert Overnight Research

CRITICAL READ. Ginnie's brutal verdict (5.5/10 — NOT READY). Cross-examination of every claim, contradictions with Babigon, 10 judge questions, Week 1 fix plan (+17 points).

Babigon's unified synthesis: Top 10 data sources, Top 10 papers/tools, 4-phase integration roadmap (260 hrs), pitch deck recommendations, SWOT analysis.

Round 1: 91 potential data sources discovered — open + paid + academic. Cast the widest net across Zenodo, NASA, ESA, USGS, BGS, commercial satellites, insurance databases, Thai government portals.

Round 2: Top 20 verified — URLs tested, data schemas inspected. Ranked by composite score (accessibility + predictive value + cost). All top 10 are FREE.

Round 3: Deep dive on top 5 sources. Actual data fields, feature mapping to SinkAlert, F1 gain estimates, sample queries, integration pseudocode.

Round 4: Step-by-step integration order. Cost breakdown (everything free for research). Time estimates. Expected F1 improvement per source (+0.02-0.08 each).

Round 1: 19 key papers with DOIs. Sinkhole prediction, InSAR+ML, pavement CV, groundwater-subsidence, karst mapping, physics-informed NN, graph neural networks, foundation models.

Round 2: 24 production-ready tools. GitHub repos (stars, activity, license), HuggingFace models, QGIS plugins, SAR processors. Includes qgis-mcp (161★) evaluation.

Round 3: Benchmark against 3 best published models. Where SinkAlert wins (multi-modal fusion) and falls short (no early warning, no temporal validation). 3 weakest links identified.

Round 4: Top 10 actionable improvements ranked by impact. 3 novel ideas no other team will have. AWS services to add. Per-dimension pitch improvements with score estimates.

CPU training baseline: mAP@50 = 0.25 (failed). Needs SageMaker GPU on full 7,598-image dataset.