Shirley XueAll work
Product + CodeHackathonMulti-agentQwen Cloud2026

TrustFlow — agents negotiate, code enforces

Problem
European enterprises are stuck between banning AI tools and uncontrolled GDPR / EU-AI-Act exposure. In Germany a works council can turn weeks of approval into months.
Solution
Agents negotiate; a compiler — pure code, same input, same policy — signs; humans activate; a gateway enforces. No model anywhere in the enforcement path.
Result
Hackathon demo, deployed live on Alibaba Cloud. I originated the concept and co-built it with a software engineer, both of us directing AI agents.
Context

Every new AI use case in a European enterprise needs sign-off from compliance, IT, procurement, and worker representatives.

TrustFlow turns that negotiation into software. Built for the Qwen Cloud Global AI Hackathon (Track 3, Agent Society): I owned the product, design and evidence side, and my teammate led the engine implementation.

Agents can propose; only deterministic code can decide. Where that line falls:

  1. The agents emit structured proposals, never executable rules.
  2. No LLM sits in the enforcement path. PII scanning, model routing and audit logging are pure code.
  3. Humans activate. Nothing enforces until the data-protection officer (DPO) and IT sign the compiled policy.
  4. A rejection gets an explainable appeal path instead of a dead end.
System

Negotiate, compile, sign, enforce

The pipeline has one hard hand-off in the middle: creative negotiation above it, hard guarantees below it.

TrustFlow architecture — employee and governance portals feed an agent boardroom, whose proposals go through a deterministic compiler and human sign-off into an active policy enforced by an edge gateway with an audit logView full size
The system architecture, end to end
  1. 01

    Agent boardroom

    Qwen-powered specialists debate the request in structured rounds — proposals only.

  2. 02

    Deterministic compiler

    Pure code merges demands, floor-checks them against the org's red lines, validates the schema, hashes and signs.

  3. 03

    Human sign-off

    DPO and IT sign to activate the compiled policy.

  4. 04

    Edge gateway

    PII scanning, model routing and audit logging on every inference.

Boardroom

Five specialists, structured disagreement

The demo models a German fintech's approval lanes with five agents, each with a real mandate: a workflow runner advocating the business case, corporate compliance holding GDPR and EU-AI-Act red lines, IT owning routing and data residency, procurement gating on vendor contracts, and a works-council liaison tracking the worker-representation agreement.

They debate in structured rounds (opening, lane statements, rebuttals, finals; up to 15 turns), reacting to the shared transcript rather than answering isolated prompts. Every turn returns a schema-validated envelope: stance, demands, concessions. Disagreement resolves into one of three deterministic outcomes: approved, denied, or pending an external gate.

The glass-box view makes the whole thing inspectable for judges: the debate transcript, the pipeline strip, and a live gateway panel where you can send a prompt through the compiled policy and watch an email get masked or an IBAN get blocked.

The TrustFlow glass box — five specialists' negotiation transcript on the left, the compiled-policy gateway panel on the right masking an email address in a live prompt, outcome ALLOWED with a local redaction hopView full size
The glass box · the payment-data scenario: agents negotiated sovereign routing, the gateway masks PII in a live prompt
Measured

Measured against a single agent

Why five agents instead of one well-prompted 'AI governance advisor'? Same request packet, same qwen-max model: a tool request with the vendor's data-processing agreement still unsigned.

The single agent conditionally approved, with audit controls only, and never mentioned the unsigned contract. The specialist boardroom caught it in round one: procurement vetoed, the debate explored workarounds, and the compiled outcome stayed DENIED. The comparison is committed to the repo and reproducible with one command.

Side-by-side baseline: one generic agent conditionally approves and never surfaces the unsigned vendor DPA; the specialist boardroom denies, with procurement blocking in round oneView full size
Same input, same model · single agent vs specialist boardroom
Product

The product around the agents

Governance software fails when only lawyers can read it. The employee portal gives a plain answer: what's approved, what's blocked, and exactly which gate is missing. A compliance-readiness meter shows the signed gates (vendor contract, works-council agreement, compiled policy hash) and what's still pending; the full boardroom trace is one tab away.

An employee can appeal a denial, which re-opens the boardroom with an advocate agent arguing their case.

TrustFlow employee portal — compliance readiness at 80% waiting for human sign-off, with signed vendor DPA, works-council agreement and compiled policy hash, and the full five-agent negotiation trace belowView full size
The employee portal · readiness, gates, and the full negotiation trace
Status

Where it stands

A hackathon build, not a product. The test suite asserts five demo scenarios: happy path, works-council gate, high-risk denial, sovereign routing, vendor-contract veto. Replay mode runs the full pipeline on recorded live transcripts, so judges could see everything without an API key.