Inspiration
AutoPal grew from a frustration: small businesses juggle invoices, recurring charges, and payment reminders across tools, making it easy to miss a duplicate or a price increase. We wanted to make money leaks easier to see without making AI the financial decision-maker. A workspace brings spending signals together, explains why something may need attention, and prepares the step for the owner to review. PayPal’s payment platform makes it possible to explore a handoff from insight to action, with a person approving before money moves. AutoPal is built around one principle: clearer decisions, human control. PayPal actions stay in human hands.
What it does
AutoPal gives small businesses a workspace to review expenses and prepare safer next steps. Users import a CSV, add expenses, inspect duplicate candidates and unusual charges, and review evidence for each risk flag. They check selectable-text invoices, draft payment reminders, compare recurring costs, estimate cash flow from data, and calculate equal or custom group splits. PayPal Sandbox authentication was tested; payment actions require human approval before capture. The AI workflow is server-side, but it produced no live result in this demo because the OpenAI account lacked quota. Sample records are labeled. The demo labels sample data and unavailable live actions.
How we built it
We built AutoPal as a lightweight web app with a static HTML, CSS, and JavaScript interface and a Cloudflare Worker backend. The interface follows a small-business workflow: review spending, inspect risk signals, check invoice text, prepare reminders, and keep payment decisions visible. CSV imports, manual expenses, selectable-text PDF extraction, risk rules, cash-flow summaries, and equal or custom group splits run locally in the browser. Sample records are clearly labeled. The Worker contains server-side integrations for the OpenAI API and PayPal Sandbox; secrets never enter the browser bundle. OpenAI output is expected to follow a validated structured format, but the API returned AI_QUOTA_UNAVAILABLE during testing, so no live AI analysis was produced. PayPal Sandbox OAuth succeeded, but the provider timed out in the earlier full-flow test; we did not create, approve, or capture a payment. Payment actions remain gated behind human approval.We added a responsive dark interface, original motion visuals, and a short hero video with pause controls and reduced-motion fallback. The app passed 54 automated tests and mobile browser checks. It is a functional prototype, not a production financial service. We also documented setup steps, provider limits, demo flow, and the safeguards judges should verify before testing the prototype.
Challenges we ran into
Two integration hurdles shaped our prototype. The OpenAI API key was recognized, but the account returned AI_QUOTA_UNAVAILABLE, so we couldn’t demonstrate a live AI-generated insight. PayPal Sandbox authentication succeeded in one test, but checkout was blocked by a timeout; we did not create or capture an order. We also found that merchant transaction history and actions require per-user authorization. Because the prototype has no sign-in or rate limiting, we kept sensitive routes closed and marked sample data clearly instead of exposing account information or pretending unsupported flows worked during evaluation.
Accomplishments that we're proud of
We’re proud that AutoPal moved beyond slides into a runnable 12-module prototype with a path from spend review to human-reviewed payment actions. Users can import CSVs, add expenses, inspect duplicate and unusual-spend flags, extract text from selectable-text PDFs, draft reminders, estimate cash flow, and split group costs. We built a responsive, motion-rich interface with a narrated product walkthrough and an installable Android package, while keeping API credentials out of client code. The project passes 54 automated tests, and browser checks cover mobile layout and key workflows. We also verified PayPal Sandbox OAuth without creating an order or moving money. When AI quota was unavailable, we surfaced the limitation instead of fabricating a result—a trust boundary we’re proud to have kept.
What we learned
We learned that financial automation only earns trust when it explains its evidence and leaves the final decision with the owner. Visuals alone are not enough: judges need to see a working path, know which data is real, and understand what happens when a provider fails. We also learned to separate local analysis from account-level actions. CSV review, manual expenses, selectable-text PDF extraction, and split calculations can run locally, while PayPal merchant data requires user authorization. Our Sandbox OAuth check succeeded, but a checkout timed out; OpenAI returned a quota error. Instead of inventing success, we labeled those limits and kept payments gated. Next, we’d add authenticated accounts and reliable provider access before expanding automation.
What's next for Autopal
Next, we’d make AutoPal safe and reliable for real business accounts. First, we’d add sign-in, per-user PayPal authorization, and rate limits so each owner sees only their own data. We’d restore live AI with a funded, monitored provider and keep deterministic checks as a fallback. Then we’d connect supported PayPal transaction, invoice, subscription, and dispute APIs, adding verified webhooks and audit trails where permissions allow. We’d test duplicate detection and cash-flow insights with varied, consented data, measure false alarms, and improve the review experience. Every payment, cancellation, or submission would remain behind a clear human approval step. Finally, we’d publish setup instructions, an open-source repository, and a short end-to-end demo.
