QuoteX turns any seller product or ambiguous multilingual buyer request into a ready-to-review cross-border offer. Qwen Cloud plans with typed function tools; deterministic code owns catalog facts, pricing, routing, and policy; persistent memory improves later decisions; and a mandatory human checkpoint prevents publication or sending without approval.
Built for Track 4: Autopilot Agent in the Global AI Hackathon Series with Qwen Cloud.
Export sales teams do not receive clean forms. They receive short emails, forwarded messages, vague references such as “the same board,” contradictory product requests, deadlines, freight ceilings, and payment expectations. A wrong answer can lose margin, promise unavailable stock, or damage a customer relationship.
QuoteX automates the repetitive work while keeping the commercial decision accountable.
- Real Qwen tool agent: Qwen3.7 chooses six strict function tools in a bounded loop: structure request, retrieve memory, match catalog, select shipping, calculate quote, and enforce policy. Every tool call, deterministic result, planner turn, latency, and audit digest is visible in Agent evidence.
- Conversational product intake: Add a product turns voice or text into editable listing facts across electronics, fashion, home, beauty, collectibles, sports, industrial goods, and resale categories. Qwen asks only for missing required details; Tablestore persists validated records and private OSS keeps product photos in cloud mode.
- Real voice intake: browser audio is sent to Qwen ASR, with quantities, dates, prices, and product codes preserved in the transcript. Browser speech recognition is a labeled compatibility fallback.
- Customer Voice Workspace: voice, text, camera, and image context share one live Qwen3.7 session. The workspace surfaces quote facts, delivery, approval state, model activity, recovery, and optional transcript instead of hiding work behind a microphone animation. Qwen Voice Design creates the QuoteX brand voice, and the matching Qwen3-TTS-VD model synthesizes it. Streaming WAV headers are normalized for browser playback, and transient signed-asset downloads use bounded retry.
- Non-trivial tool workflow: Qwen plans, but trusted code executes memory relevance, phrase-polarity catalog matching, freight optimization, margin-safe pricing, risk assessment, and the approval policy. The model cannot set a price or bypass a gate.
- Production-safe degradation: missing, timed-out, malformed, or quota-limited Qwen calls trigger bounded guardrail completion. Recovery is labeled and persisted; it is never presented as live model success.
- Durable run ledger: the last 200 agent runs are stored in Tablestore in cloud mode, with model, planner turns, skill coverage, latency, human gate, and a SHA-256 evidence digest. SQLite remains the zero-setup local adapter.
- Persistent learning: approved outcomes are written to a versioned browser store, recalled on future RFQs, capped per customer, and expired after 365 days.
- Human control: every commercial offer stops at a human approval gate. Model text never directly triggers an outbound send.
- Multimodal output: Qwen3.7 grounds a campaign brief in the uploaded photo, a Wan/Qwen image router creates the commercial edit, and HappyHorse turns the resulting frame into a five-second product video. Image URLs are downloaded immediately; video task state and provider provenance remain visible.
- Marketplace adapters: one verified seller record produces validation-first Amazon, eBay, and Alibaba.com draft payloads with platform-specific condition mappings, title limits, marketplace IDs, missing fields, warnings, and JSON export. Publishing remains disabled until OAuth and human approval exist.
- Measured baseline:
npm run benchmarkexecutes 100 checked-in workflows and currently reports 100% product selection, quote arithmetic, approval-gate enforcement, and no-model fallback completion. It deliberately makes no unmeasured human-time claim. - Measured architecture advantage: a live six-case adversarial evaluation gave the governed Qwen tool agent 42/42 checks versus 28/42 for the same Qwen3.7 model producing the decision in one prompt, a 33.3-point lift. The fixtures, scorer, result, and limitations are checked in.
- Executable Alibaba Cloud stack: official SDKs provision Tablestore, private OSS, SLS, and a least-privilege RAM role; an AMD64 publisher captures the immutable ACR digest; the FC3 SDK creates or updates the Custom Container and protected HTTP entry point.
- Live Alibaba runtime: the ACR-free build is deployed on Function Compute in Japan (Tokyo). Its public health check reports the real FC runtime and configured Qwen services; paid AI routes remain protected by a private access token.
The primary submission film is the 1080p frame-accurate Remotion judge cut. It uses real application clips, continuous camera choreography, kinetic type, and synchronized typewriter, interface, shutter, impact, and whoosh sound design. The original evidence-led recording remains available as a slower technical backup.
- Start with the 500-scarf Ulaanbaatar-to-Berlin buyer request, its product photo, and four recalled buyer requirements.
- Open Agent evidence to show all six Qwen-selected skills, trusted results, audit digest, and the 42/42 versus 28/42 adversarial comparison.
- Open Offer to show recomputed pricing and the locked human checkpoint.
- Open Campaign to compare the source photo with the generated image, then show validation-first Amazon, eBay, and Alibaba.com drafts.
- Show the prepared customer answer with Qwen3.7 and Qwen Voice Design both complete.
- Switch to Resilient demo to prove the same six tools run with zero model turns and the send gate still closed.
- End on the architecture diagram.
The upload-ready local artifacts, exact timecodes, narration, and QA evidence are in docs/CINEMATIC_DEMO.md and docs/DEMO_SCRIPT.md.
flowchart LR
U[Voice · text · photo] --> API[Validated Node boundary]
API --> P[Qwen3.7 bounded planner]
P -->|typed function calls| T[Six verified business tools]
T --> MEM[(Customer memory)]
T --> CAT[(Catalog + inventory)]
T --> QUOTE[Freight + margin-safe quote]
T --> POLICY[Risk policy]
POLICY --> GATE{Human approval}
GATE -->|approve| OUT[Offer + learned outcome]
GATE -->|hold| STOP[No outbound action]
P --> LEDGER[(Tablestore run ledger + audit digest)]
API --> OSS[(Private OSS product media)]
OUT --> MEDIA[Qwen vision · image edit · HappyHorse]
OUT --> MARKET[Amazon · eBay · Alibaba drafts]
The Node service owns secrets, live Qwen planning, verified tool execution, and the durable audit ledger. The browser owns the editable workbench and can run the same deterministic safety path when live inference is disabled. See docs/ARCHITECTURE.md for data flow, failure modes, memory lifecycle, and threat model.
The architecture diagram also ships as a judge-ready SVG, PNG, and editable Excalidraw file.
| Role | Default | Reason |
|---|---|---|
| Tool planning + RFQ extraction | qwen3.7-plus |
Strong multilingual function calling; strict schemas, missing-tool repair, and a four-turn ceiling keep agency bounded |
| Customer answers | qwen3.7-plus |
Grounded, policy-bounded customer support |
| Product-photo understanding | qwen3.7-plus |
Flagship multimodal grounding and structured output |
| Voice transcription | qwen3-asr-flash |
Full-recording multilingual transcription through the compatible API |
| Voice identity | qwen-voice-design |
Creates a warm, conversational QuoteX brand voice from a natural-language description |
| Spoken response | qwen3-tts-vd-2026-01-26 |
Synthesizes the designed voice; target model is required to match the design request |
| Primary image edit | wan2.7-image-pro |
Recommended starting point for brand control and image editing |
| Image fallback | qwen-image-2.0-pro |
Strong semantic adherence and negative-prompt control |
| Product video | happyhorse-1.0-i2v |
Animates the uploaded or AI-edited first frame through a traceable async task |
This routing follows the official model selection, structured output, speech recognition, voice design, speech synthesis, vision, and image editing guides. Environment variables can override each role when a workspace exposes a different model set.
Requirements: Node.js 22.15 or newer. npm install provides the Alibaba Tablestore and OSS runtime adapters; local mode uses Node's built-in SQLite module.
npm install
cp .env.example .env
npm startOpen http://127.0.0.1:4173.
The deterministic demo works without credentials. To prove live Qwen usage, set at least:
QWEN_API_KEY=
QWEN_BASE_URL=https://dashscope-intl.aliyuncs.com/compatible-mode/v1
QWEN_MODEL=qwen3.7-plus
QWEN_AGENT_MODEL=qwen3.7-plus
QWEN_VISION_MODEL=qwen3.7-plus
QWEN_ASR_MODEL=qwen3-asr-flash
QWEN_VOICE_DESIGN_MODEL=qwen-voice-design
QWEN_VOICE_DESIGN_TARGET_MODEL=qwen3-tts-vd-2026-01-26
QWEN_TTS_MODEL=qwen3-tts-vd-2026-01-26
QWEN_TTS_VOICE=
QWEN_VIDEO_MODEL=happyhorse-1.0-i2vUse a separate regional image key when required:
QWEN_IMAGE_API_KEY=
QWEN_IMAGE_MODEL=wan2.7-image-pro
QWEN_IMAGE_FALLBACK_MODEL=qwen-image-2.0-pro
QWEN_IMAGE_BASE_URL=https://your_image_workspace.ap-southeast-1.maas.aliyuncs.com/api/v1The key and every endpoint must belong to the same region. An international key works with dashscope-intl.aliyuncs.com; a workspace key works with that workspace's dedicated compatible and DashScope endpoints. They are not interchangeable. Enable both qwen-voice-design and qwen3-tts-vd-2026-01-26: the first creates a custom voice, while the second synthesizes speech with it. The first successful design is cached in .runtime/; leave QWEN_TTS_VOICE empty to create it automatically. Qwen may apply its documented voice-design quota or charge. Secrets are read only on the server and never returned to the browser.
Seller listings default to .runtime/quotex.sqlite for local development. Cloud mode uses Alibaba Tablestore for listing metadata and agent-run evidence plus private OSS for original product photos. Function Compute accesses both through short-lived credentials from a least-privilege RAM execution role.
npm test
npm run typecheck
npm run build
npm run benchmark
npm run evaluate
npm run verify:submission
npm run deploy:plan
npm run probe:qwen
npm run probe:qwen-image
npm run probe:happyhorse
curl http://127.0.0.1:4173/api/healthprobe:qwen-image creates one small image-edit request. probe:happyhorse creates the minimum three-second 720P video and is billed by Qwen per generated second.
Use npm run evaluate -- --live to compare the live governed Qwen tool agent with the live one-prompt Qwen baseline. The test suite covers Qwen tool-call contracts, model-proposed quantity conflicts, bounded recovery, prompt-injection resistance, memory-query isolation, agent-run persistence, product intake and SQLite CRUD, phrase-polarity catalog matching, marketplace adapters, RFQ parsing, Qwen ASR, customer-safe grounding, Qwen Voice Design/TTS reuse, HappyHorse tasks, pricing, risk escalation, memory, approval, image grounding, model failover, and generated-asset persistence.
| Mode | What happens | Judge-visible proof |
|---|---|---|
| Live Qwen | Qwen chooses typed business tools in a bounded server loop | Evidence shows model, turns, each tool call/result, host, latency, tokens, and digest |
| Resilient demo | Runs the same trusted business core without the planner | Every guarded completion is labeled; quote logic and human gate still work |
The fallback is a reliability feature, not a simulated Qwen success. QuoteX never labels a deterministic result as a live model call.
QuoteX has two executable Function Compute deployment paths:
code-package: an ACR-free 2.85 MB ZIP that runs on Function Compute's built-in Node.js 20 executable. It is the fastest judge-demo route and uses explicitly non-durable memory storage unless Alibaba Tablestore and OSS are configured.custom-container: an immutable AMD64 ACR image with Tablestore, private OSS, SLS, and a least-privilege RAM role for the durable production architecture.
- Browser application: Open QuoteX
- Alibaba backend health:
GET /api/health - Alibaba API base:
https://quotex-utopilot-vybltedhtp.ap-northeast-1.fcapp.run - Region and runtime:
ap-northeast-1,custom.debian10, built-in Node.js 20 - Deployed backend source: commit
3c80dab - Browser bridge source: commit
4e2f960 - Machine-readable proof: docs/alibaba-deployment-evidence.json
Open the GitHub Pages URL in a browser. Alibaba Cloud adds Content-Disposition: attachment to responses from its generated Function Compute domain, so that domain is intentionally used as the API backend rather than the browser entry point. The static browser client calls the same Function Compute deployment through a strict CORS allowlist. Live paid AI endpoints require the private judge fragment and return 401 without it; the fragment is removed from the address bar and retained only for the current browser session.
The latest browser smoke test returned HTML without a download header, recorded zero console or network failures, reached Alibaba listings with 200, and completed a live qwen3.7-plus run with all six skills and the human approval gate. This judge deployment honestly reports storage.provider: "memory" and durable: false; the checked-in Tablestore, OSS, SLS, and RAM path is the durable production deployment.
Build and inspect the ACR-free request:
npm run deploy:prepare
npm run deploy:package
npm run deploy:plan
npm run deploy:fcSet ALIBABA_FC_DEPLOYMENT_MODE=code and QUOTEX_STORAGE_PROVIDER=memory for that demo route. The full apply sequence remains npm run provision:alibaba, npm run image:publish, and npm run deploy:fc; it provisions managed storage and logging, captures an immutable ACR digest, then creates or updates Function Compute and its HTTP trigger. Both dry runs redact every Qwen key and the private demo token.
Devpost specifies a repository code-file link as the required Alibaba Cloud deployment proof. Use server/alibaba-fc-deployment.ts: it constructs and executes the official FC3 CreateFunction/UpdateFunction and HTTP-trigger requests. server/alibaba-cloud-infrastructure.ts provisions Tablestore, OSS, SLS, and RAM, while server/qwen-tool-orchestrator.ts shows the live Qwen Cloud runtime boundary. The separate deployment evidence binds the running endpoint and ZIP digest to the exact deployed commit.
The code-proof requirement is distinct from additional runtime evidence. The Function Compute API URL, artifact digest, health response, browser headers, CORS checks, and authenticated Qwen smoke test were recorded only after deployment succeeded. Follow docs/DEPLOYMENT.md for the evidence contract and both reproducible deployment paths.
src/
main.ts Typed workbench and human checkpoint
rfq-engine.ts Catalog, memory, pricing, freight, risk, and audit logic
marketplace-adapters.ts Amazon, eBay, and Alibaba validation drafts
memory-store.ts Versioned cross-session memory with expiry and bounds
qwen-client.ts Live/fallback mode and browser-to-server boundary
types.ts Shared domain and API contracts
server/
qwen-tool-orchestrator.ts Bounded Qwen function-calling planner
agent-evaluation.ts Fair governed-versus-single-prompt evaluator
alibaba-cloud-infrastructure.ts Tablestore, OSS, SLS, and RAM provisioning
alibaba-storage.ts Durable Tablestore and OSS repositories
alibaba-fc-deployment.ts Idempotent FC3 function and trigger deployment
agent-run-store.ts SQLite agent evidence ledger
listing-store.ts Validated SQLite seller listing and photo repository
qwen-parser.ts Hardened Qwen extraction and trace capture
qwen-asr.ts Validated Qwen speech-to-text boundary
customer-agent.ts Customer-safe quote support and escalation
qwen-tts.ts Qwen Voice Design, cached voice reuse, and audio persistence
happyhorse-video.ts Async product-video submission and task polling
marketing-asset.ts Qwen vision brief + Wan/Qwen image router
config.ts Environment and endpoint configuration
tools/
serve.ts Typed secure HTTP server
benchmark-agent.ts Repeatable deterministic quality baseline
evaluate-agent.ts Six-case live/offline adversarial evaluation
provision-alibaba-cloud.ts Idempotent managed-resource provisioning
publish-alibaba-image.ts Secret-safe AMD64 ACR publish and digest capture
deploy-alibaba-fc.ts Secret-safe FC3 create/update command
tests/ TypeScript engine and multimodal integration tests
dist/ Generated JavaScript; ignored by Git
docs/ Architecture, deployment, demo, and rubric evidence
