QuotePilot AI

Final approved email, complete activity log, and system architecture
Clarification state with activity log and human-in-the-loop architecture
Extracted requirements and five clarification questions
Approved $11,800 quote with requirements, email, and activity log
Deterministic quote prepared while the availability check runs
Initial dashboard: every agent stage waiting before the request runs
Live Qwen workflow begins with Intake Analysis running
Qwen detects missing details and pauses for clarification
Mandatory Human Approval checkpoint before simulated email and CRM

Inspiration

Service businesses lose hours turning incomplete customer requests into accurate quotes. The work spans requirement analysis, clarification, pricing, availability, approval, email, and CRM systems. Fully autonomous quoting is risky because a model-generated price or premature customer message can damage trust. We wanted an agent that moves quickly while keeping business rules and final authority under control.

What it does

QuotePilot AI turns an unstructured service request into structured requirements, clarification questions, a deterministic quote, an availability check, and a customer-ready summary. A visible six-stage workflow stops whenever required information is missing and stops again at Human Approval. No simulated email or CRM action completes without explicit approval.

The demo workflow produces an $11,800 website and marketing automation quote after one clarification round.

How we built it

Qwen Cloud performs request understanding, structured requirement extraction, missing-information detection, clarification generation, allow-listed package recommendation, and customer-facing summary drafting.
Deterministic business tools control catalogue lookup, pricing, discounts, tax handling, and demo availability. Qwen never sets prices or availability.
Zod validates requests and all structured model output.
Next.js, React, TypeScript, Node.js, and Tailwind CSS power the responsive UI and server routes.
Alibaba Cloud Function Compute hosts the production application in Singapore using a Debian 12 custom runtime with Node.js 20.

Architecture

Customer Request → Next.js UI → Alibaba Cloud Function Compute → Qwen Cloud Agent → Deterministic Business Tools → Human Approval → Simulated Email and CRM

This separation lets Qwen reason flexibly about language while deterministic code protects business commitments. The approval boundary is visible in both workflow state and the runtime activity log.

Challenges

The central challenge was balancing agent autonomy with predictable business behavior. We separated probabilistic language work from deterministic pricing, validated package IDs against the requested service, rejected malformed provider output, and made stopping conditions explicit. We also prepared a self-contained Next.js standalone deployment for Alibaba Cloud Function Compute and secured the Qwen configuration in server-side environment variables.

Accomplishments that we're proud of

Live Qwen intake and clarification workflow
Deterministic quoting protected from model-generated prices
Explicit approval and rejection paths
Controlled provider-failure behavior with no fabricated fallback
Responsive dashboard and safe health endpoint
Public HTTPS deployment on Alibaba Cloud Function Compute

What we learned

Reliable agents need narrow tools, validated contracts, explicit stopping conditions, and observable state. Giving the model less authority produced a stronger product: Qwen can reason flexibly while code protects the commitments that matter to the business.

Limitations

QuotePilot AI is a hackathon prototype. Email and CRM actions are simulated, and the service catalogue and availability are demo datasets. It does not claim real customer usage, production revenue, or live CRM/email integrations.

What's next

Real CRM, calendar, email delivery, and e-signature integrations; authenticated workspaces; persistent audit history; catalogue administration; multilingual intake; retrieval from business policies; evaluations, retries, rate limits, and queued workflows.