Threshold

Inspiration Climate disasters rarely arrive without warning; the ocean gives signals weeks, sometimes months in advance. Sea surface temperatures climb. Dissolved oxygen plummets. Coral bleaching alerts spike. Yet humanitarian funding consistently lags behind these signals, chasing headlines instead of data. We were inspired by a visceral statistic that early climate intervention costs 10x less than post-crisis recovery. We wanted to build the tool that closes that window, a Bloomberg Terminal for climate tipping points, routing capital before the cameras arrive.

What it does Threshold is a proactive climate crisis intelligence platform that calculates a "Days to Threshold" countdown for eight critical ocean ecosystems (Great Barrier Reef, Coral Triangle, Mekong Delta, Arabian Sea, Bay of Bengal, California Current, Gulf of Mexico, Baltic Sea). It

Ingests 20+ live data sources from NOAA, NASA Ocean Color, Coral Reef Watch, Scripps Keeling Curve CO₂, World Bank, OCHA financial tracking, and more Runs a BiLSTM + multi-head attention neural network trained on 8 oceanographic time series channels to produce a tipping-point proximity score on [0, 10] Surfaces a global triage queue ordered by urgency, exposing intervention windows before mainstream media coverage Opens a direct funding loop where Stripe Connect and Solana/USDC payouts route capital immediately to KYC-verified NGOs, bypassing institutional delays Provides counterfactual cost-benefit analysis showing if you fund now vs. in 6 months, what does the delta in ecosystem recovery cost look like

How we built it: We used a full-stack architecture designed for speed and scientific rigor

Layer Stack Frontend React 18 + Vite, Globe.gl (3D globe), D3.js, Tailwind CSS, Stripe SDK Backend FastAPI (async Python), SQLAlchemy 2.0, Snowflake (SQLite fallback) ML/AI PyTorch (BiLSTM), XGBoost, Prophet, Google Gemini 1.5 Data Pandas ETL pipeline, 20+ REST APIs Blockchain Solana (Anchor/Rust), Web3.js, USDC DevOps Docker Compose, AWS CDK/Lambda Threshold takes in 8 normalized feature channels per timestep: ΔSST, DO, DHW, BAL CO2, ΔCHl, NO3-, CI and outputs a scalar proximity score s in [0,10] where values above 7 trigger high-urgency alerts. Risk scoring weights are grounded in IPCC AR6 and EPA/NOAA standards. Gemini 1.5 Pro augments the pipeline with NLP-based urgency scoring from GDELT and ReliefWeb crisis reporting.

Challenges we ran into API rate limits and quota exhaustion meant constant juggling of fallbacks, and every service needed a mock/stub layer so the app stayed functional during demos. Snowflake cold-start latency in a hackathon environment pushed us to build a SQLite fallback that mirrors the full schema. Solana devnet instability forced us to wrap wallet and program health checks defensively and add an is_mock_mode flag so the funding flow could be demonstrated without a live chain Model training time was a real constraint since Threshold needed real oceanographic sequence. Fusing heterogeneous signals was the hardest data problem, merging ocean physics data (SST anomalies in C°), geopolitical conflict indices, and financial tracking numbers into a single normalized feature vector required careful domain research

Accomplishments that we're proud of: An end-to-end pipeline spanning raw ocean sensor data through ML inference, through interactive globe to one-click NGO funding, all wired together Threshold, a purpose-built BiLSTM that produces interpretable, scientifically-grounded tipping-point scores rather than black-box outputs A counterfactual engine that puts a dollar figure on delay, showing decision-makers exactly what waiting costs A clean Solana USDC payout flow with KYC-protected recipient verification, not a mock checkout, but actual on-chain settlement logic

What we learned Ocean science is rich with open data. NOAA, NASA, Coral Reef Watch, and Scripps all provide \ accessible APIs once you understand their conventions Attention mechanisms make a real difference on multivariate time series with asynchronous signals. The multi-head attention layer over BiLSTM hidden states improved our tipping-point recall noticeably over a vanilla LSTM baseline Blockchain-native payments solve a real humanitarian problem. Traditional wire transfers to NGOs in affected regions take days, while on-chain USDC settlement is nearly instant Fallback engineering is as important as the happy path. In a hackathon (and in production), the system that keeps running wins What's next for Threshold Expand to 30+ ecosystems, including Arctic sea ice extent, Amazon basin moisture, and Sahel vegetation indices Live Snowflake refresh pipeline running on 6-hour cadence against real NOAA/NASA feeds Foundation model fine-tuning to replace Threshold's hand-crafted features with a climate foundation model (e.g., ClimaX) pre-trained on ERA5 reanalysis data NGO dashboard where verified organizations get a real-time view of incoming USDC, disbursement receipts, and impact reporting tied back to the tipping-point scores that triggered their funding Policy API to expose Days-to-Threshold scores via a public REST endpoint so governments, reinsurers, and multilateral development banks can integrate the signal into their own decision systems

Built With

• amazon-web-services
• brev.dev
• gemeniapi
• node.env
• orthogonal
• react
• scripps