Inspiration
Humanitarian organisations depend on reliable payment processes to deliver financial assistance. However, when payments fail, become delayed, are duplicated or require correction, resolving those exceptions can involve reviewing multiple records, checking supporting evidence and coordinating decisions across different people.
We wanted to explore a different approach:
What if AI could help investigate payment exceptions, understand supporting evidence and prepare recovery recommendations, while keeping financial decisions under human control?
This question led to ReliefRelay, an AI-assisted humanitarian payment recovery platform designed to make payment exception investigation more structured, transparent and accountable.
What it does
ReliefRelay brings payment case investigation, evidence review, recovery proposals, human approval and payment tracking into one operational workflow.
Its key capabilities include:
- AI-powered investigation: Analyses payment case information and supporting evidence to generate structured investigation findings.
- Evidence-grounded recommendations: Connects investigation findings to available case evidence rather than relying on unsupported conclusions.
- Recovery proposals: Helps prepare and validate proposed payment recovery actions.
- Human approval: Separates AI recommendations from financial authorisation. An authorised human must approve a proposal before a payment action proceeds.
- PayPal payment integration: Integrates with PayPal Payouts for payment execution and status tracking.
- Auditable decision history: Maintains a hash-linked record of important workflow events, decisions and actions.
The central principle is simple: AI assists with investigation, but humans remain accountable for financial authorisation.
How we built it
ReliefRelay combines an AI investigation layer, a structured backend, payment infrastructure and a web-based operational interface.
Technology stack:
- Frontend: React and Vite
- Backend: Node.js and Express
- Database: SQLite
- AI: OpenAI API
- Payment infrastructure: PayPal Payouts API
- Deployment: Cloudflare Workers
The workflow begins with a payment exception case and its supporting information. The AI investigation layer analyses the available evidence and produces structured findings. These findings support the preparation of a recovery proposal, which then passes through validation and a separate human approval process.
Once authorised, the payment workflow can proceed through the PayPal integration. Important decisions and workflow events are recorded in the audit history.
PayPal integration and testing
During development, we implemented and tested the PayPal Payouts integration using a private PayPal Sandbox environment.
We successfully completed one GBP 12.50 PayPal Sandbox payout and verified the successful PayPal batch and payout item responses.
The publicly accessible judge demo uses synthetic data and a local PayPal stub. It does not connect to PayPal Live or PayPal Sandbox and does not send payments. The private Sandbox integration and public demonstration are separate environments.
Challenges we ran into
One of our biggest challenges was ensuring that AI assistance did not become autonomous financial authorisation.
In a payment recovery workflow, an AI-generated recommendation should never automatically become an authorised financial action. We addressed this by separating investigation, proposal preparation, validation, human approval and payment execution into distinct stages.
Another challenge was maintaining traceability. It is not enough to know that an action happened; the system should also preserve the sequence of decisions and supporting workflow events.
We also had to balance a realistic public demonstration with payment safety. We chose synthetic data and a local PayPal stub for the public demo, allowing judges to explore the application without triggering financial operations.
Accomplishments that we're proud of
- Built a working end-to-end AI-assisted payment recovery prototype.
- Implemented evidence-grounded AI investigation within the case workflow.
- Developed separate recovery proposal, validation and human approval stages.
- Integrated the PayPal Payouts API and successfully completed a private GBP 12.50 PayPal Sandbox payout.
- Implemented a hash-linked audit history for workflow accountability.
- Deployed a publicly accessible demonstration on Cloudflare Workers.
- Published the source code under the AGPL-3.0 licence so the project can be inspected and developed further.
We are particularly proud of bringing AI investigation, human financial oversight and payment infrastructure together in one coherent workflow.
What we learned
Building ReliefRelay reinforced that applying AI to financial workflows requires much more than generating useful responses.
AI outputs need supporting evidence. Financial decisions need accountable human approval. Payment actions require explicit controls. Important decisions must remain traceable.
We also learned how to integrate AI-assisted investigation with payment infrastructure while maintaining a clear separation between public demonstrations and private Sandbox testing.
Most importantly, we learned that trust and accountability need to be designed into the workflow from the beginning, not added afterwards.
What's next for ReliefRelay
ReliefRelay is currently a working prototype. Our next priorities are:
- Expand testing across more payment exception and recovery scenarios.
- Improve validation of failed, returned, unclaimed and cancelled payment workflows.
- Extend PayPal integration and webhook testing.
- Strengthen audit verification and operational controls.
- Evaluate the platform with humanitarian organisations and potential pilot partners.
- Measure whether the workflow can reduce investigation effort and improve payment recovery transparency.
Our longer-term goal is to make payment exception handling more evidence-driven, accountable and efficient for organisations delivering humanitarian assistance.
Project links
- Live demo: https://reliefrelay.nostologia.com
- Source code: https://github.com/pradeepannagarasu/RELIEFRELAY
- Demo video: https://youtu.be/ozs4sxMhZzw
- Licence: GNU Affero General Public License v3.0 (AGPL-3.0)
Built With
- api
- artificial
- cloudflare
- express.js
- financial
- human-in-the-loop
- intelligence
- javascript
- node.js
- openai
- payment
- payouts
- paypal
- processing
- react
- rest
- sqlite
- technology
- vite
- workers