Tresora

Tresora — The Smart CFO Assistant

Crypto treasury finally runs like a modern finance team.

The Core Problem

Crypto treasury operations are slow, manual, and dangerously fragmented.

Paying a single vendor invoice today means:

Chasing context: DMing people on Telegram for addresses and confirmations

Manual math: Cross-checking spreadsheets for budgets, balances, and gas

High-risk execution: Manually swapping, bridging, and transferring funds — the #1 source of treasury losses

Result: A task that should take 5 minutes turns into 3–5 days of coordination, risk, and human error.

The Solution

Tresora doesn’t replace the CFO — it replaces the busywork.

We introduce a Multiplayer AI Treasury, where specialized AI agents work in parallel to prepare a complete, verifiable “Treasury Plan.”

The human CFO’s role is reduced to what actually matters: 👉 Final review and signature.

No private keys. No blind execution. Full control.

The Multiplayer AI Architecture

Tresora uses a role-based agent system, where each AI has a clear responsibility and permission boundary.

1. The Requester — Intake & Context

Role: Staff proxy Does:

Listens to chat commands

Parses unstructured data (PDFs, messages)

Example:

Extracts amount, currency (USDC / EURC), and recipient address directly from a PDF invoice

1. The Sentinel — Risk, Compliance & Intelligence

Role: The brain + security layer

Checks in real time:

Budget: “Do we have sufficient USDC on Arc Testnet?”

Compliance: “Is this address sanctioned?” (Chainalysis / Etherscan)

Market Risk: “Are gas fees abnormally high?”

Flags anything risky for human review — no silent failures.

1. The Executor — On-Chain Builder

Role: Transaction architect

Does:

Constructs the full on-chain execution path

Never touches private keys

Queues transactions inside Circle Developer-Controlled Wallets with strict policies

Execution only happens after explicit human approval.

Live Workflow (Demo)

Trigger: User uploads a PDF invoice in chat:

“Pay this vendor.”

Agent Coordination (The Magic):

Requester: “Extracted payment request: 5,000 USDC → 0x742…”

Sentinel: “Insufficient USDC on Arc Testnet. Liquidity found on Sepolia.” “Risk checks passed. Gas fees slightly elevated — flagged.”

Executor: “Treasury Plan generated:

Swap ETH → USDC on Sepolia (Uniswap v4)

Bridge USDC to Arc via Circle CCTP

Transfer USDC to vendor”

Resolution: The CFO opens the dashboard, reviews the single unified Treasury Plan, and clicks:

👉 Sign & Execute

The exact sequence runs — nothing more, nothing less.

Why This Wins

Parallel AI agents instead of linear automation

Human-in-the-loop by design, not as an afterthought

Safer than manual execution, faster than centralized tooling

Built for real treasuries, not demos

Technical Stack

AI: Gemini — fast reasoning + PDF understanding

Swaps: Uniswap v4 (Sepolia)

Bridging: Circle CCTP (USDC-native, trust-minimised)

Wallets: Circle Developer-Controlled Wallets (policy-based execution)

Chain: Arc Testnet — instant, low-cost, compliance-friendly

How it's made This project is a sophisticated multi-agent system built on Next.js 14, orchestrating complex DeFi operations through specialised AI agents powered by LangGraph.

The AI Brain (LangGraph & Gemini) We architected the agent workflow using LangChain's LangGraph. This allowed us to build a stateful, cyclical graph where three specialised agents (Requester, Sentinel, Executor) pass messages and coordinate tasks autonomously.

Agent Orchestration: LangGraph manages the conversation state and hands off tasks between agents. For example, the "Requester" parses the invoice and passes the structured data to the "Sentinel" node for validation. The "Sentinel" (Security Layer): Before any transaction is built, this agent calls the Etherscan API to perform a real-time audit of the receiving wallet. It checks for past suspicious activity, contract verification status, and token holding patterns to ensure the treasury isn't sending funds to a malicious actor. The DeFi Engine (Uniswap V4 & Circle) We chained together multiple protocols to create a seamless "Pay Invoice" experience, abstracting away the complexity of cross-chain liquidity.

Uniswap V4 (Sepolia): We integrated the Uniswap V4 SDK to programmatically swap ETH for USDC when the treasury is short on stablecoins. The V4Quoter fetches real-time quotes, and the UniversalRouter executes the swap within a single atomic transaction flow. Circle Programmable Wallets & CCTP: Wallets: We utilised Circle's Developer-Controlled Wallets to hold treasury funds, allowing us to execute policy-based transactions via API rather than exposing private keys. CCTP (Cross-Chain Transfer Protocol): When the "Sentinel" detects the vendor is on Arc Testnet but funds are on Sepolia, it automatically orchestrates a burn-and-mint bridge transaction via CCTP. This moves USDC across chains without slippage or traditional bridge risks. Notable "Hack": The Stateful Planner The most challenging part was handling asynchronous blockchain operations (Swap → Wait → Bridge → Wait → Transfer) within a chat interface.

Our Solution: We built a custom "Treasury Plan" state machine inside the LangGraph workflow. The AI doesn't just execute; it drafts a multi-step execution plan. The frontend polls this plan while the server recursively executes each step, making a complex, multi-chain operation feel like a single, atomic "One-Click Pay" action for the user.