Inspiration

A payout can succeed at PayPal while an internal operational record still says “pending.” That leaves an operator with a difficult question: what actually happened, and what is safe to do next?

Sending another payout could pay the recipient twice. Correcting the record without checking the evidence could hide a real problem.

RecoveryOS explores how AI can help investigate this discrepancy while keeping the recovery decision grounded in verified provider evidence, explicit policy, and operator approval.

What it does

RecoveryOS is an interactive incident lab built around a genuine PayPal Sandbox payout.

Each visitor starts an isolated demo session. After PayPal confirms that the specific payout item has reached SUCCESS, the demo deliberately skips the corresponding internal record update. This injected fault is clearly labeled, and reconciliation automatically opens an incident.

The operator can then run an AI investigation across five evidence sources:

  • The incident
  • The internal payout record
  • The PayPal payout item
  • The event history
  • The recovery policy

The investigation returns a diagnosis, cited evidence, and a proposed action. The operator reviews the proposal before approving a record correction.

At approval time, the backend independently reads the PayPal item again and checks the recovery policy. If verification succeeds, it corrects the internal record, resolves the incident, and appends an immutable audit event.

The correction changes only the internal operational record. It never sends another payout.

How it is built

The frontend uses React, TypeScript, and AG Grid to present the incident journey, evidence, investigation, and audit history.

A Go backend with PostgreSQL manages session isolation, payout creation limits, recovery policy, and atomic persistence of the correction and its audit event.

The PayPal Sandbox Payouts API provides genuine payout creation and item-level status evidence. The OpenAI Responses API supports investigation through five read-only tools and structured output. The AI has no tool for creating payouts, approving proposals, or modifying records.

Challenges

A central challenge was separating a useful AI recommendation from permission to change operational state. RecoveryOS addresses this through backend validation and fresh provider verification at approval time.

Another challenge was making an external-service demo repeatable. Sandbox payouts can take time, and AI services can be unavailable. RecoveryOS provides two clearly labeled fallback paths: a pre-created genuine Sandbox payout and a deterministic diagnosis. Both retain the same approval checks.

Preserving evidence was equally important. The interface distinguishes provider observations, injected events, AI conclusions, and fallback conclusions so the operator can understand where each finding came from.

Accomplishments

RecoveryOS connects genuine Sandbox evidence to a complete, inspectable recovery workflow: discrepancy detection, investigation, operator review, verified correction, and audit history.

Each visitor can explore that workflow without registering and without changing another visitor’s internal record or incident.

What I learned

The project reinforced that reliable AI-assisted recovery depends on the surrounding system: precise evidence references, constrained tools, explicit policy, and transactional state changes.

It also highlighted why payout-item status matters. A batch-level result alone does not establish the state of the specific payout being investigated.

What’s next

The next step is to expand the incident scenarios and evaluate investigation quality across more ambiguous evidence, while preserving the same boundaries around operator approval, provider verification, and auditable record changes.

Share this project:

Updates

Submission history