Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

OweMyGod: a fine‑tuned 8B model reads Jakarta debt excuses right 92% of the time (keyword rules: 60%)

Description and use

OweMyGod employs a fine‑tuned 8‑B Qwen3 model to interpret informal excuses and schedule debt reminders in Jakarta, reducing cognitive load and preserving friendships.

Problem

Friends who lend money in Jakarta struggle to chase repayments without damaging relationships.

Outcome

Fine‑tuned Qwen3‑8B achieved 92% exact‑date accuracy and 100% kind‑right accuracy on 25 hand‑written Jakarta debt excuses, outperforming keyword rules (60% accuracy) and the untrained model (68%). The app reduced early reminders from 14% to 8% and prevented over‑nagging by limiting reminders to one per 3 days.

Workflow

  • 1. User logs debts and partial payments in the app.
  • 2. User copies a friend’s excuse from WhatsApp into the app.
  • 3. The LLM parses the excuse and outputs a "kind" and "when" JSON.
  • 4. Python code converts the "kind" into an exact date based on user’s timezone and settings.
  • 5. A schedule engine decides the next reminder time and drafts a tone‑appropriate WhatsApp message.
  • 6. The app sends a single 09:00 push notification when a debt is due or a promise is made for that day.
  • 7. The user reviews the drafted message and sends it via WhatsApp.
  • 8. The app tracks reminders, stops after three firm reminders without response, and suggests in‑person discussion.

Evidence

The model reads Jakarta slang; code does arithmetic.

Source

Fine‑tuned v2 (shipped) 92% exact date, 8% too early, 100% kind right.

Source

The app never sends two reminders within 3 days, and after three firm ones with no answer it tells you to talk in person and stops.

Source

  • Record date: 2026-10-24T10:11:34.465Z