Takken Drill for Business

AI judges who really studied. HR approves with one click. PayPal pays employees directly.

Comment

・Inspiration In Japan, every real-estate office must employ at least one licensed Takken holder for every five staff, and only licensed staff may explain key contract terms to customers. So companies pay for their employees' exam fees and courses. But in 2025 only 18.7% of 245,462 candidates passed, and many company-funded candidates barely studied. HR has no fair way to tell genuine effort from clicking through, and no simple way to reward it.

・What it does Takken Drill for Business changes the deal between a company and its employees:

Employees who study genuinely get the exam fee reimbursed, pass or fail. Employees who pass get a bonus. Employees who did not study pay the fee themselves. Employees study in a full drill app (4-choice drills, true/false review, mock exams, a study schedule). Every answer is recorded.

When HR presses Judge with AI, Claude reads each employee's last 8 weeks of accuracy and seconds per question and returns an effort score, a pass probability, flags with evidence, and a proposed award with written reasons in Japanese and English. It catches patterns that simple time-based rules miss:

memorized answers (88% accuracy in 9 seconds per question), clicking through (35%, close to random guessing), a last-minute spike after 7 quiet weeks. HR reviews the proposals in an AG Grid table, edits amounts or sends a follow-up message, and approves with one click. The money goes straight from the company's own PayPal Business account to the employee's PayPal through the Payouts API. The app never holds the money.

・How we built it Frontend: React + TypeScript + Vite, AG Grid for the HR review table, hosted on Cloudflare. Backend: Supabase (Postgres with row-level security per company, Auth, Edge Functions in Deno). AI: Claude (claude-haiku-4-5) with structured JSON output: per-employee judgement plus a company-wide report. Only pseudonymous references are sent; no names. PayPal (sandbox): Payouts API from the company's own PayPal app, with sender_batch_id to prevent double payment. Webhooks registered automatically on the company's app and verified with verify-webhook-signature. Log in with PayPal so payouts go to the employee's verified PayPal ID, not a typed email. Subscriptions API + JS SDK button for the company's monthly fee for the app. Guard rails: the server clamps every AI amount to the company's award rules; if the AI fails, the award waits for a human; employees see the AI's reasons only after an admin confirms. Cost control: the AI runs only on demand, identical input reuses the stored result (hash cache), and there is a daily cap per company.

・Challenges we ran into Making the AI judgement fair and explainable: rules on study time alone are easy to game, so the AI judges the combination of accuracy and speed and the change over time, and must give evidence for every flag. Keeping money safe: the AI only proposes. A person approves, the server clamps amounts, and PayPal batch IDs and webhook handling make each payout happen exactly once. Letting each company pay from its own PayPal account (client secret stored AES-GCM encrypted) instead of the app holding funds.

・Accomplishments that we're proud of An end-to-end flow that works: study → AI judgement with reasons → one-click approval → real PayPal sandbox payout → webhook updates the status. The employee can see why the AI decided what it did, which makes the system feel fair rather than surveillant.

・What we learned AI is most useful here as a careful reviewer that writes its reasons, with a person making the final call. PayPal Payouts + Log in with PayPal make paying many people simple without the app ever touching the money.

・What's next for Takken Drill for Business Pilot with real-estate companies in Japan for the October 2027 exam. Payroll integration as an alternative payout method (a CSV export already exists, since such awards may be taxable as salary). The same model for other company-funded certifications.

Built during the hackathon

Takken Drill for Business was built from scratch during the submission period. Work started on October 4, 2026, three days after the period opened on October 1. Everything on the company side is new:

  • AI judgement: the Claude integration with structured output, the per-employee judgement and the company report, the hash cache and the daily cap.
  • HR approval flow: the admin dashboard with the AG Grid review table, editing amounts, follow-up messages, self-pay, and award rules clamped on the server.
  • PayPal integration: Payouts from each company's own PayPal app, webhook verification and handling, Log in with PayPal for employees, and Subscriptions for the monthly fee.
  • Backend: the Supabase database with row-level security per company, consent before any answer is recorded, and all Edge Functions.
  • Hosting: the live demo with fictional employees, refreshed daily.

The one existing piece is the employee study app. It is adapted from my own consumer study site, takken-drill.com, which existed before the hackathon. For this project I removed its ads, membership plans and login. I then connected it to the new system, so every answer is reported for the AI judgement, and added the "Award" tab, where employees see their award, its status, the AI's reasons and their PayPal account.

The public repository's first commit (October 6) contains the work from October 4–6; all later changes are in the commit history.

Built With

Share this project:

Updates

Submission history