AI-managed crypto vault that lets anyone ride Polymarket alpha. 7-agent autonomous risk engine on Sui mainnet, prediction-market signals + ZK-attested decisions.
ZkVanguard is an early-stage TypeScript project in the AI payments / x402 ecosystem. It currently has 4 GitHub stars and 2 forks.
An AI-managed USDC vault on Sui — deposit once, autonomous agents allocate & hedge for you.
Prediction markets print billions per month in alpha-bearing signal. Riding it consistently needs bots, capital, and 24/7 attention — table stakes for hedge funds, impossible for retail. ZkVanguard collapses that workflow into a one-click USDC vault: 7 AI agents, BlueFin V2 perp hedging, 3-way reconciliation, 8-gate autonomy defense, ZK-STARK-attested at scale. Live on Sui mainnet.
Contents — What your USDC does · Verify in 60s · Safety · Status · Live metrics · Roadmap · Built by · How it works · Revenue · FAQ · Quickstart · Docs
Fees: 50 bps annual mgmt + 10% performance. Both charged by the Move contract itself and routed to an MSafe multisig — public, auditable, no operator custody.
# No clone required — hits live production endpoints
curl -s https://www.zkvanguard.xyz/api/health/production | jq
curl -s https://www.zkvanguard.xyz/api/predictions/per-asset | jq
# With clone — canonical "is the pool in profit?" script
bun run scripts/analyze-pool-pnl.ts
bun run scripts/check-hedge-signal-alignment.ts
Or inspect the pool object directly on Suiscan:
0x107292a69eea2f6eaf4a4e4727ee25d747b04c1985441b138933f0ef33f7b7260xe814e0948e29d9c10b73a0e6fb23c9997ccc373bed223657ab65ff544742fb3aShipped July 2026 after a real drawdown revealed passive-only defenses. Each gate defends a specific failure mode; the full stack is verified by an integration test that replays the historical scenario and asserts max NAV loss ≤ 15%.
| Gate | Defends against |
|---|---|
| PortfolioDriver | Existing spot never unwound when profit-lock fires — corrective actions actively reshape the balance sheet |
| Fill verifier | BlueFin "silent-reject" — orders returning orderHash but never landing on the exchange |
| Hedgeability spot-cap | At small NAV, perp minQty makes hedging impossible — spot cap for that asset forced to 0 |
| Symmetric sell trigger | Rebalance was one-sided (bought on 65% conviction, never sold on 65% opposing) — now symmetric |
| Stale-hedge detector | Positions > 7d old with ≥ 2 signal flips force-close on contradiction |
| Signal-flip drift-close | On direction flip, both perp and spot legs unwind — not just perps |
| AI regret weighting | Position size shrinks after losing streaks, recovers on wins — prevents AI-euphoria buying tops |
| Alert response loop | 3 KILL alerts/hr auto-shrinks spot; 24h profit-lock auto-unwinds; phantom hedge rate > 1% halts trader |
Verify: bun jest test/integration/pool-drawdown-defense.test.ts (10/10 green).
Always-on structural guards: 2-of-3 agent consensus on trades > $100K · 10% peak-NAV drawdown halt · circuit breaker after 3 failures · 3-way reconciliation (Move ↔ BlueFin ↔ Postgres) · OFAC geo-block (KP/IR/SY/CU/RU/BY) · strict NAV-oracle mode (deposits/withdrawals revert when cron oracle > 2h stale).
Live on Sui mainnet (v0.2.0, deployed 2026-06-12). Pre-external-audit, TVL deliberately capped at $10K by contract. Cap lifts after external audit closes — the constraint is intentional operational proof, not a TVL claim.
15 internal audit phases completed pre-mainnet. Engine has been running autonomously since June 2026 with continuous on-chain NAV snapshots; every production incident to date has been caught, remediated, and documented in the deploy record and docs/DEPLOY_RUNBOOK.md.
Snapshot pulled from the Aiven Postgres replica against live Sui mainnet state. Rerun bun run scripts/analyze-pool-pnl.ts any time.
| Metric | Value |
|---|---|
| Days running since first NAV snapshot | 46+ (from 2026-05-29; live-computed) |
| NAV snapshots recorded | 2,200+ (≈48/day, matches 30-min cron cadence) |
| Hedges executed lifetime | 214 across BTC / ETH / SUI / SOL PERPs |
| Active crons with heartbeats | 13 — see /api/health/production |
| Active members | 3 (limited by $10K TVL cap) |
| Lifetime USDC deposits | ~$38 (rerun scripts/analyze-pool-pnl.ts for live number) |
| External auditors engaged | Pending (SUI Foundation grant Tranche 1 deliverable) |
Small absolute numbers by design — the $10K cap is enforced by the contract. Operating metrics prove the engine works; audit + cap-lift unlock scale.
| Quarter | Milestone |
|---|---|
| Q3 2026 | External audit close · TVL cap ratchet $10K → $100K · Founding-100 points program live |
| Q4 2026 | TVL cap ratchet to $1M · Institutional tier live (custody attestations via rwa_custody_attestor.move) · first EVM chain expansion |
| Q1 2027 | TVL cap ratchet to $10M · Enterprise white-label API · TGE (utility token, governance + fee-share) |
Cap ratchets are contract-gated via admin_set_tvl_cap and gated on each milestone's success criteria (audit pass, incident-free operating window, TVL sustained). Not aspirational — each unlock is a specific contract call after a specific evidence bundle.
Multi-chain posture. SUI is the lead chain by design and gets new features first. Cronos, Oasis, Hedera, and Sepolia contracts are compiled, tested, and configured in hardhat.config.cjs; deployment triggers on institutional demand from that chain, not a race.
Ashish Regmi (@HarveReg) — CS, Cryptography + AI majors. Formerly senior engineer at multiple Fortune 500 companies. Multi-chain hackathon winner across EVM, Aptos, and ICP. Upstream contributor to the Tether WDK EVM wallet (accepted PR). Solo builder — open to core contributors and institutional partners.
Contact: [email protected] · Telegram @anstemple
flowchart LR
PM["Polymarket + Manifold + Delphi<br/>(prediction markets)"] --> SF
BF["BlueFin funding rates"] --> SF
CDC["Crypto.com 24h momentum"] --> SF
SF["Signal fusion +<br/>synthetic-STRONG layer"] --> AG
AG["7 AI agents<br/>2-of-3 consensus<br/>SafeExecutionGuard"] --> PD
AG --> ZK["ZK-STARK attestation<br/>(trades > $1M)"]
PD["PortfolioDriver<br/>+ 8 autonomy gates"] --> VA["USDC vault<br/>BTC · ETH · SUI"]
PD --> PERP["BlueFin V2 perps<br/>+ 7k aggregator (6 DEXes)"]
Live today: 50 bps annual mgmt + 10% performance fee on vault deposits — charged automatically by community_pool_usdc.move, routed to FeeManagerCap on MSafe.
Post-audit (all mapped to shipped code in lib/config/pricing.ts):
Token: utility-first, designed not launched. Governance over fee parameters, staking gates discounted vault fees. TGE targeted Month 9–12 post-audit.
Node 18+, Bun, Python 3.11+, PostgreSQL.
git clone https://github.com/ZkVanguard/ZkVanguard.git && cd ZkVanguard
bun install --legacy-peer-deps
# Terminal 1 — ZK-STARK prover
python -m pip install -r zkp/requirements.txt
python zkp/api/server.py
# Terminal 2 — Next.js
bun run dev
# Pre-commit
bun run typecheck && bun run lint && bun jest
Full architecture, env conventions, BlueFin invariants, and reconciliation topology: see docs/ARCHITECTURE.md, docs/DEPLOY_RUNBOOK.md, and docs/SUI_DEPLOYMENT.md.
How do I deposit and withdraw? Connect a Sui wallet at zkvanguard.xyz, approve USDC, deposit. Withdrawals are always open — the Move contract computes your USDC share of NAV (idle pool + admin wallet + BlueFin margin) and pays out atomically. Strict NAV-oracle mode adds a safety: if the cron oracle attestation is > 2h stale, both deposits and withdrawals revert to prevent bad pricing.
What if the AI is wrong? Every trade > $100K needs 2-of-3 agent consensus. The 8-gate autonomy defense catches most failure modes automatically. Drawdown > 10% from peak NAV auto-halts new hedges. AI regret weighting shrinks position sizes after losing streaks. Beyond that, you can withdraw at any time.
What if the operator disappears?
Withdrawals are non-custodial — the Move contract holds pool USDC and computes payouts against the on-chain state. Fees route to a MSafe multisig, not a hot wallet. Off-chain BlueFin margin is the one operational dependency; the sui-hedge-reconcile cron sweeps it back to the pool address every hour, and close_hedge funds-verify (via AgentCap) prevents drain scenarios.
Why is TVL capped at $10K?
Deliberate. The cap is enforced by the Move contract itself (admin_set_tvl_cap) and lifts only after external audit closes. Operating publicly at a bounded size is proof — same code, same crons, same defense stack, capital-constrained until reviewers sign off.
How is this different from Ethena / GMX / Aave GHO? Ethena is delta-neutral USDe backed by ETH shorts. GMX is a perp DEX. Aave GHO is over-collateralized. ZkVanguard is none of these — it's an AI-directional vault that uses prediction markets as its alpha source and perps only to hedge, not to yield-farm. Closest analog is a Yearn v3 strategy vault, but with signal-driven allocation instead of yield harvesting.
What if BlueFin has an outage?
bluefin-health cron runs a 3-strike de-risk (close-all reduceOnly on venue degradation). bluefin-db-reconcile (every 15 min) and sui-hedge-reconcile (hourly) sweep any orphaned positions when the venue recovers. The vault continues to accept deposits/withdrawals against on-chain NAV throughout.
Responsible disclosure: report security issues privately to [email protected]. Do not file public issues for active vulnerabilities.
Built on Sui, BlueFin V2, Polymarket, Manifold, Crypto.com, Aiven, Upstash, Vercel, and the Tether WDK ecosystem.
The agent-native LLM router for autonomous agents. 55+ models (8 free), <1ms local routing, USDC payments on Base & Solana via x402.
A payments protocol for the internet. Built on HTTP.
A local-first AI agent with persistent memory, emotional intelligence, and a peer-to-peer skills economy.
The trust layer for agent-to-agent commerce — natural-language mandates, ERC-7710 delegated permissions, x402 payments, escrow, and dispute resolution as one open, catch-all Agent Skill / Claude Code plugin.
Self-healing infrastructure for AI agent payments. 90.3% auto-recovery.
The AI agent with a wallet — spends USDC autonomously to get real work done. Apache-2.0, TypeScript.