Qwen wrote its own app: why QuotePilot still needs a human approval gate

QuotePilot is a hackathon project that turns cross-border B2B inquiry emails into approved, bilingual (EN/中文) price quotes. What makes it notable isn't just that Qwen models power it at runtime — much of the app itself was written by Qwen models dispatched through a small harness, with the developer reviewing and accepting or rejecting each piece. The entire build cost under $1 of a $40 hackathon credit.

The architecture rests on one firm rule: the LLM never does arithmetic and never writes legal text. Pricing is computed in code using Python Decimal, and legal clauses are fixed, pre-written bilingual text. Models only map customer intent to catalog items and draft cover copy — they never touch the money math.

The build surfaced real engineering lessons: a cloud provider's file-download behavior forced an architecture change, asking a code model to re-emit a large file caused it to silently drop a feature, and an adversarial multi-agent review round caught 11 real bugs. Frame-by-frame review of the demo video itself even revealed a mismatch between pre- and post-approval quote data.

The key design decision was making the system pause for exactly one human approval rather than running fully autonomously. For a document that becomes a contract, that human gate — knowing which outputs to trust and which to check line by line — is the actual engineering takeaway.