How Qwen 3.8 Max Turns Plain English Into On-Chain Guardrails

Sentinel — AI Policy Assistant, built for the Monad Metropolis hackathon (Alibaba Cloud / Qwen 3.8 Max bounty)

The problem: nobody knows how to configure a guardrail

Sentinel puts hard limits on AI agents before they can move money on-chain: spending limits, recipient whitelists, time-locks, and circuit breaks. The contracts enforce them. But there was a UX gap — the people who most need guardrails (solo developers running trading agents) know what they want ("don't let my agent lose more than X"), but not how to translate that intent into smart-contract parameters.

What does "1 MON per transaction" become? What's a sane daily limit? How long should a time-lock be? Before Qwen, users had to hand-fill six numeric fields and hope.

Qwen 3.8 Max closes that gap. We integrated it as the AI Policy Assistant: you type your intent in plain language, Qwen parses it into exact on-chain policy parameters, and the policy is created — a real transaction on Monad Testnet.

The integration

The assistant lives in frontend/src/components/AiAssistant.tsx and calls the DashScope API (OpenAI-compatible mode) with the qwen-max model:

POST https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions
model: "qwen-max"
temperature: 0.3

The system prompt tells Qwen it is a policy-configuration assistant for an on-chain guardrail system, and constrains it to respond only with valid JSON matching a fixed schema:

{
  "maxSpendPerTx": "<number in MON>",
  "maxSpendPerPeriod": "<number in MON>",
  "periodDuration": "<seconds>",
  "timeLockDuration": "<seconds>",
  "timeLockThreshold": "<number in MON>",
  "circuitBreakThreshold": "<number in MON>",
  "whitelist": ["<contract names or addresses>"],
  "explanation": "<brief explanation of what was configured>"
}

The response is parsed, mapped to a PolicyParams object, and pushed straight into the policy-creation form. The user reviews the prefilled numbers and submits — the result is a real on-chain policy in the SentinelRegistry/SentinelGuard contracts.

What Qwen does for us

1. Natural language → policy parameters. "Limit my agent to 1 MON per transaction, whitelist only Uniswap, circuit-break if it loses 50 MON in an hour" comes back as exact numbers: maxSpendPerTx: 1, whitelist: [...], circuitBreakThreshold: 50, plus a human-readable explanation.

2. Smart defaults from intent. "I want my agent to trade safely on DeFi" yields a conservative configuration rather than forcing the user to invent numbers.

3. Plain-English explanations. Every generated policy ships with an explanation field so users understand what each parameter does — without reading Solidity.

4. Monad-aware context. The system prompt anchors defaults in MON amounts and typical DeFi transaction sizes on Monad's fast blocks (1 MON/tx, 10 MON/day, 1-hour time-lock above 5 MON, circuit break at 50 MON).

Graceful degradation

Production engineering matters in a hackathon demo. The assistant has a two-layer fallback:

• If the DashScope call fails (network, key, rate limit), a local regex parser extracts MON amounts, time-locks, and whitelist names from the same sentence.
• If VITE_QWEN_API_KEY isn't set at all, the local parser runs directly.

The UI never breaks, and the user always gets a working policy — Qwen makes it intelligent, the fallback makes it reliable.

Why Qwen specifically

• Structured-output discipline. qwen-max follows the JSON-schema instruction reliably at temperature: 0.3 — we get parseable policy objects, not prose.
• The compatible-mode endpoint means no bespoke SDK: one fetch, standard OpenAI-style payloads, trivially portable.
• Value per token is the right shape for this task. Policy parsing is short-context, high-precision work — exactly what a flagship model should be spent on.

Try it

1. Open https://sentinel-monad.vercel.app
2. Go to the Create Policy page and open the AI Policy Assistant
3. Type: "Strict mode: 0.5 MON per tx, whitelist only Uniswap, auto-pause at 2 MON"
4. Watch the form fill itself in — then submit to create a real on-chain policy on Monad Testnet

Repo: https://github.com/ubongn/sentinel
Bounty: Alibaba Cloud — Best Builds with Qwen 3.8 Max