ClaimBack My Money

ClaimBack My Money - Your loan records, made clear. A free self-help agent for people who lent money to a friend and never got it back. It reads your own chat screenshots, transfer receipts and IOU with Meta models, has you confirm every fact, and fills Korean court-document drafts that you file yourself. Every number comes from your confirmed facts and deterministic code, never from the AI. THE PROBLEM In Korea, loans between friends are often made over chat with a bank transfer and maybe a handwritten IOU. When the money is not repaid, the cheapest route is to send a certified demand letter (내용증명) or file a payment order (지급명령) yourself. Turning scattered screenshots into a dated, numbered, redacted evidence set and consistent document text is where people give up or pay for help. WHAT IT DOES Upload your chats, transfer receipts and IOU (or load the synthetic sample case). Role agents work through the files, and you check their work at two checkpoints: - Reader (Muse Spark 1.2) reads each image with bounding boxes; long screenshots are read in overlapping tiles. - Vision (SAM 3.1) finds chat bubbles, profile pictures, stamps and signature boxes. - Fact organizer (Muse Spark 1.3) proposes each value (people, amount, dates, repayments) tied to the exact reading it came from, flags conflicts and asks questions. - Privacy (rules) burns account, ID and phone numbers out of the image pixels and flags any mask it cannot confirm. - Checkpoint 1: you confirm or correct every value. - Calculator (deterministic code) splits repayments between interest and principal and applies statutory rates. - Drafter (Muse Spark 1.3) and Verifier (rules) loop: the drafter writes fact sentences with slots only ({{principal}}, {{quote:EX1:item10}}); the verifier rejects literal numbers or names, computed values posing as agreements, facts without evidence and quotes the evidence does not contain, and asks for a rewrite. - Draft approval: you approve or edit each sentence. - Packager renders the demand-letter PDF, the payment-order text to paste into the e-court form, and the exhibit bundle with "갑 제N호증" marks and an evidence list. A final check re-reads every PDF against your confirmed values. SAFETY BY DESIGN It is not legal advice: no win predictions, no "you can claim" judgments, no legal labels on what people said. The AI writes slots, not values. Redaction is burned into pixels before any PDF is built. All sample evidence is synthetic and marked SAMPLE. BUILT DURING THE HACKATHON Code through 2026-11-02 was written by the owner during the hackathon. From 2026-11-03, Muse Code implemented most milestones while Claude Code wrote the specs and task briefs and reviewed each one; the final fixes were made by Claude Code with two reviewer agents cross-checking. The screen's visual style (layout, colours, font) is adapted from the owner's earlier project; the case flow, the agents, the screen logic and the English screen were built during the hackathon. Task briefs, reports, reviews and a ledger of every live Meta API call are in the repo. TRY IT The demo video was recorded on live Meta Model API calls (34 calls, $0.19). The repo runs locally in replay mode with no API key (README, Quick start). The screen is in Korean and English; the court documents stay Korean because they are filed in Korean courts.