Inspiration
One income payment has to cover several obligations. “10% for tithe, 20% to mummy, 10% for rent” sounds simple, but turning that instruction into a trustworthy plan takes more than a chatbot. We wanted AI to help with the planning while keeping amounts and payment approval under human control.
What it does
Ravasend Plan runs an allocation agent and an independent safety agent concurrently through Mozaik. Real local Qwen2.5 inference turns the instruction into a proposed budget. Deterministic integer-cent checks preserve the total and reserve, reject invalid percentages, and catch omitted or changed explicit allocations. The user reviews and confirms before PayPal sandbox checkout.
PayPal OAuth and Orders v2 create a single sandbox merchant order. The buyer approves through PayPal’s hosted sandbox checkout, then the app captures only the order bound to the confirmed plan. This is not a multi-beneficiary payout integration: family, rent and tithe are budget labels, not actual recipients.
How we built it
New isolated work started 30 October 2026 using TypeScript, Express, Mozaik and local Ollama/Qwen2.5 0.5B. Two fire-and-forget runLoop calls share request-scoped state with a coordinator observer. Both answers must arrive before validation. Agents have no payment tools. Sandbox calls use idempotency IDs and order binding. No production Ravasend code, customer records, live credentials or real funds are included.
Verification
The public repository passes TypeScript and all 14 tests, including 100 rounding/reconciliation cases and concurrent-runtime checks. Real local-model runs show both inference starts before either completes. A real PayPal sandbox create/approve/capture run returned COMPLETED for order 7T797569G0808353K. A synthetic US sandbox merchant is a test fixture only; it does not change Ravasend’s Nigerian incorporation or imply live US availability.
Challenges and learning
The small local model initially omitted an allocation and raised a false safety warning. We improved the prompts and added a deterministic consistency guard rather than trusting the model’s arithmetic or completeness. Embedded checkout also needed a hosted-approval fallback. Offline fixture mode remains available for orchestration tests and is explicitly labelled—not presented as AI.
What is next
Better clarification for ambiguous instructions, broader evaluation of model errors, and persistent reconciliation before any production use. The accompanying captioned walkthrough uses genuine captures from the working prototype and completed sandbox test.
Log in or sign up for Devpost to join the conversation.