The autonomous financial circuit breaker for DeFi. Detects systemic risk (liquidity crises, depegs) through a quantitative model, and automatically evacuates users' funds into a safe stablecoin — before the pool collapses.
▶ Live demo: umbra-ten-iota.vercel.app · Risk terminal: /terminal
🏆 ETH Global Lisbon 2026 — Tracks: The Graph · 0G · (Uniswap v3 execution)
📄 Research paper — full risk-score methodology: PDF · Word. Single source (paper/content.py), regenerable via python paper/build_paper.py (Word) and python paper/build_pdf.py (PDF).
🎤 Pitch deck (math, on-brand 16:9): paper/Umbra_Slides.pdf — python paper/build_slides.py.
When a DeFi pool goes into crisis (depeg, liquidity flight), the average user realizes it too late: by the time they understand what's happening, liquidity has evaporated and exit slippage has exploded. The losses of March 2023 (USDC depeg), Terra/UST, stETH… run into the billions.
An on-chain circuit breaker that continuously monitors a Risk Score (CBRI, 0→100) and triggers an emergency evacuation as soon as the optimal threshold τ* is crossed.
- Quant model (CBRI) — 3 signals aggregated via Noisy-OR: liquidity-flight speed, pool imbalance, price divergence.
- Optimal threshold
τ*— proven by backtesting on real crashes: the exact point that maximizes net funds saved. - Aligned business model — a success fee charged only on the loss avoided. We only earn if the user wins.
| Layer | Tech | Role |
|---|---|---|
| Data 🏆 | The Graph | Historical + real-time Uniswap v3 pool data (tick-level swaps, liquidity, price) |
| AI / Infra 🏆 | 0G | Decentralized storage & traceability of the model and risk scores |
| Execution | Uniswap v3 | Direct on-chain evacuation swap (QuoterV2 + SwapRouter02, no API) |
🏆 = targeted tracks. Uniswap = data infra + execution (we are not applying to their track).
Backtest vs. production rigor: the backtest simulates execution against historical on-chain liquidity (the pool's real physics via The Graph); live execution reads an on-chain Uniswap QuoterV2 quote then swaps via SwapRouter02. We price the past with physics, and execute the present at the pool's real price.
umbra/
├── quant-backtest/ # Python — the brain (CBRI model + backtesting)
└── live-execution/ # TypeScript — the muscle (Uniswap execution + 0G traceability)
Replayed on 48,066 real swaps ($5.57B of volume) extracted via The Graph, for a $1M USDC position.
| Optimal τ* | 66 / 100 (optimal plateau [10–66], 0 false positives in calm markets) |
| Trigger | Mar 10, 14:10 UTC — 17h before the bottom, USDC still at $1.0000 |
| Depeg bottom (no action) | $0.8726 → position worth $872,595 |
| 💰 Funds saved | $126,900 (12.7%) — 10% success fee = $12,690 |
The cost of waiting (every CBRI point you wait = money lost):
| τ threshold | Exit | Price | Slippage | Saved |
|---|---|---|---|---|
| 10–66 (τ*) | Mar 10, 14:10 | $1.0000 | 5 bps | $126.9k |
| 67–72 | Mar 11, 00:15 | $0.9892 | 53 bps | $111.4k |
| 73–98 | Mar 11, 01:00 | $0.9726 | 91 bps | $91.1k |
| 99 | Mar 11, 03:00 | $0.9371 | 657 bps | $2.9k |
The active liquidity of the USDC/USDT pool collapses by ×23,000,000 during the crash → exit slippage explodes. Exiting at τ* = near free; waiting for confirmation = a liquidity wall.
3 sub-signals normalized by a sigmoid, aggregated via a weighted Noisy-OR (a breaker trips if a single signal turns red):
CBRI = 100 · (1 − ∏ᵢ (1 − wᵢ·sᵢ)) sᵢ = σ(αᵢ·(xᵢ − thresholdᵢ))
| Signal | Measure | Role | Source |
|---|---|---|---|
| Liquidity flight | LP withdrawal speed (mints−burns / TVL·h) | early warning | swaps + mints/burns |
| Order-Flow Imbalance | flow unidirectionality | diagnostic (w=0 here, non-discriminant) | swaps |
| Divergence / depeg | |1 − USDC price| | confirmation | USDC/USDT stable pool |
A circuit breaker you have to trust blindly is worthless. Each risk score and the exact model that produced it are frozen into an attestation, whose 0G Merkle root hash is published on 0G decentralized storage. Anyone can re-download the artifact and verify the hash → auditable and tamper-proof scoring.
cd live-execution && npm run publish0g
# 🌳 0G Storage root hash: 0xc9926d168f786c07df854fa4774528396abee05b57a13fe260f0a64a1d47f90bThe 0G root hash is computed locally (no wallet required). The actual upload to the 0G testnet only needs a funded wallet (faucet).
cp .env.example .env # THEGRAPH_API_KEY (required) · RPC_URL (optional, public default)
cd quant-backtest && pip install -r requirements.txt
python thegraph_client.py # ① extract the crash data (The Graph)
python features.py # ② compute the CBRI
python backtest.py # ③ backtest τ* + figures
cd ../live-execution && npm install
npm run publish0g # ④ anchor model+scores on 0G (traceability)
npm run demo # ⑤ replay the depeg → the breaker evacuates via Uniswap v329 unit tests covering the core logic (model, slippage, τ* selection, execution):
# Quant (Python) — 21 tests: sigmoid, v3 price, Noisy-OR, slippage, τ*, funds saved
cd quant-backtest && pip install -r requirements-dev.txt && python -m pytest
# Execution (TS) — 8 tests: position conversion, slippage bps, minOut, feed parsing
cd live-execution && npm testFull-stack E2E — chains the entire pipeline (data → CBRI → backtest τ* → 0G anchoring → Uniswap evacuation) and verifies every step (14 checks):
./e2e.sh # 🟢 Full chain operational — dress rehearsal before the demoPlus 7 quant e2e tests on real data (pytest -m e2e) validating the pitch numbers (τ*=66, ~$127k saved, bottom $0.8726).
Hackathon MVP — public architecture assumed (no privacy/MEV layer in this scope).