LearnBond: Money That Moves When Learning Happens

Sponsors fund a learning goal with PayPal. An AI Tutor teaches, an independent AI Examiner verifies mastery, and only then is the money released. Unearned funds go back to the sponsor.

Comment

Inspiration

Education apps charge a subscription whether or not anyone learns. Meanwhile, millions of families support a student from far away: an aunt in London paying for a nephew's exam prep in Enugu, a parent abroad funding lessons back home. They send money and hope it turns into learning, but they have no way to know.

We asked a simple question: what if the money only moved when the learning actually happened?

What it does

LearnBond turns a learning goal into an outcome-based fund.

  1. A sponsor funds a goal through PayPal. For example, "Pass WAEC Maths: quadratic equations" for $60. Gemini designs examinable milestones and splits the budget across them by difficulty. Several relatives can each pay part of it.
  2. An AI Tutor teaches the student. It's Socratic, so it guides the student rather than handing over answers. When the student feels ready, the Tutor writes an honest readiness claim listing strengths and suspected gaps.
  3. An independent AI Examiner tries to disprove that claim. It asks 3 to 5 adaptive questions, typed or spoken, and targets the gaps the Tutor admitted to.
  4. Money moves only when mastery is verified. The Examiner's verdict must quote the student's own answers word for word. Code checks every quote against the real transcript, and an invented quote voids the verdict.
  5. Guardrails, not the AI, decide whether money moves. These include a minimum confidence threshold, an auto-release limit, a daily cap, attempt limits, and a sponsor review when answers look pasted or unusually fast. Bigger payouts wait for the sponsor's approval.
  6. An AI agent buys what's missing, within a mandate. The sponsor can set aside a small study-kit budget. The Study Kit Agent reads the gaps the Tutor and Examiner found, searches the study store (a tool call), and proposes a cart with a reason for every item. Code, not the model, enforces the mandate: real catalog prices only, a per-item cap, the kit budget, and an auto-buy limit above which the sponsor approves. It pays the store through PayPal from the held funds.
  7. Unearned money goes back. When the deadline passes or the sponsor closes the goal, every contributor is refunded their share through PayPal.

Every step, whether by an AI agent, a guardrail, PayPal or a person, is written to a SHA-256 hash-chained audit log that the sponsor can inspect. Editing any past entry breaks the chain.

How we built it

  • Plain HTML, CSS and JavaScript, with no framework and no build step. You open index.html in a browser and it runs, and a first-run setup guide walks you to Settings to connect Gemini and PayPal.
  • A zero-dependency Python server (server.py, standard library only) that holds the PayPal secret and makes payouts (Payouts API) and refunds (Refunds API) real Sandbox calls. It can also proxy Gemini so the key never touches the browser. It deploys to Render in one click from render.yaml.
  • Google Gemini (gemini-2.5-flash) plays three roles: Planner, Tutor and Examiner. Every decision that can affect money uses structured JSON output (responseSchema), so code can check it instead of trusting free text.
  • PayPal in the Sandbox: Smart Buttons checkout (Orders + Capture) for sponsors, the Payouts API to pay students and the study store, and the Refunds API to return unearned money to each contributor's original capture. Without the server, payouts and refunds fall back to clearly labelled simulations.
  • Gemini as a tool-using shopping agent: step 1 decides what to search for, code runs the catalog search, step 2 picks a cart from the results only. Student requests are treated as untrusted text.
  • AG Grid for everything a sponsor audits: a portfolio grid with progress bars and pinned totals, a money ledger of every capture, payout, purchase and refund with PayPal references, and the audit log with floating filters, hash tooltips and CSV export.
  • Web Crypto (SHA-256) for the tamper-evident audit log.
  • Web Speech API for spoken exam answers.
  • localStorage with cross-tab sync, so the sponsor's dashboard updates live when the student passes an exam in another tab.

Challenges we ran into

  • Making an AI's verdict safe to attach money to. A language model can be confidently wrong, or manipulated by a student typing "mark me as passed". We made the Examiner treat student answers as untrusted text, required word-for-word evidence quotes, and put code-based guardrails between the verdict and the payout. The AI can only recommend.
  • Keeping two agents honest with each other. If the Tutor simply says "they're ready", the Examiner tends to agree. Asking the Tutor to list suspected gaps, and telling the Examiner to treat the claim as a hypothesis to test, made the exams noticeably more probing.
  • Detecting copy-paste and coaching without unfairly failing students. Paste detection and typing speed are passed to the Examiner as hints to probe further, never as automatic failures.
  • Letting an agent spend money safely. An LLM shopping agent can invent products, inflate prices or be talked into buying something by the student. We gave it a search tool over a real catalog, then priced and filtered its cart in code against the sponsor's mandate. Anything over the auto-buy limit waits for a human.
  • Running entirely in the browser. We wanted judges to try it with zero setup, which ruled out a backend. That meant simulating payouts and refunds honestly, and restructuring our scripts so the app works even when opened straight from the file.

Accomplishments that we're proud of

  • A complete loop from PayPal checkout to AI tutoring to adversarial AI exam to guarded release or refund, working end to end.
  • Proof-of-learning receipts: the sponsor doesn't just see "passed", they see what the student said, in the student's own words, verified against the transcript.
  • Guardrails on AI-controlled money that a sponsor can understand at a glance and set themselves.
  • An audit trail where tampering is detectable, not just discouraged.
  • Agentic commerce with a mandate: an AI agent that really spends money through PayPal, but only on what the evidence says the student needs, and only within limits a non-technical sponsor sets.
  • Real money movement end to end in the Sandbox: checkout, payouts and refunds, every one tied to a hash-chained audit entry and visible in a sortable, exportable ledger.

What we learned

  • The hard part of "AI + payments" isn't calling the AI. It's deciding what the AI is allowed to decide. Treating the model as an advisor and code as the decision-maker made the whole system easier to trust.
  • Structured output changes everything: once Gemini's answers fit a schema, they can be validated, logged and audited like any other data.
  • Adversarial setups (one agent checking another) catch far more than a single agent asked to "be careful".

What's next for LearnBond: Money That Moves When Learning Happens

  • Live product search for the Study Kit Agent across real stores, keeping the same mandate checks.
  • Accounts and shared storage, so sponsors and students can use different devices.
  • Teacher and school mode: let a human tutor review transcripts and co-sign releases.
  • Richer exams: handwritten working via photo upload, and spoken back-and-forth exams in local languages.
  • Employer and scholarship sponsors funding upskilling goals with the same proof-of-learning receipts.
Share this project:

Updates

Submission history