AdultingGPT

AdultingGPT: meet Gege, the AI agent that handles adulting for you. Powered by NVIDIA Nemotron on Nebius, he tracks your money, bills and calendar and tells you what to do next

Comment

Inspiration

Adulting is mostly admin: which bill is due, whether you can afford that thing before payday, who you owe a call, what clashes on Thursday. Budgeting apps show you what you already spent. We wanted an agent that tells you what is safe to do next, and does the boring parts when you say so. AdultingGPT is built Nigeria-first (₦, Lagos defaults, Naija-style money talk) but works anywhere, and everything is optional: nothing needs a bank link, you just tell Gege what you have.

What it does

Gege is the adulting AI agent inside AdultingGPT. You talk to him (text, voice or a photo of a receipt) and he:

  • keeps the books: expenses, income, balances, loans with interest, IOUs, bills and payday;
  • works out what is actually free until payday and answers "can I afford it?" from your own numbers;
  • runs your calendar and people: events, clashes, birthdays, who you have not spoken to;
  • briefs you each day, and only acts with your approval on anything risky (every action is logged and undoable);
  • catches ideas on boards and turns an idea into a project with first tasks, grouped by stage;
  • joins your group chat: mention @Gege to split a bill, add to a shared list or set a reminder, and he proposes it for approval;
  • works from the app (iOS, Android, web) and an MCP server; chat apps like WhatsApp and Telegram are on the roadmap.

How we built it (Personal AI track)

Gege runs as an ensemble of NVIDIA Nemotron models served on Nebius Token Factory, with one model per job:

  • Nano (fast role): triages every message and does small clean-up jobs.
  • Super (agent role): tool calling, turning what you say into actions.
  • Ultra (reasoning role): careful money answers, the daily briefing and the final write-up of hard questions.
  • Vision role: reading receipts, with a fallback chain (Gemini, then DeepSeek) so Gege keeps working if a model is down.

Model ids are discovered from the Token Factory /models list rather than hard-coded, and an admin web dashboard lets us switch which model does which job, test a prompt, and see a per-reply model trace and call telemetry. Tavily powers web search ("what does a T-bill pay now?") and reading shop pages that block apps. Backend: FastAPI + SQLAlchemy; app: Expo React Native (iOS, Android, web). The cash-flow forecast, bills, calendar and insights are deterministic code; the models only choose tools and phrase answers, so numbers are traceable.

Challenges

Keeping an agent trustworthy with money: safe actions run and can be undone, risky ones become proposals, and every number comes from code rather than the model. We also built the store-compliance basics early (consent naming each AI provider, in-app privacy policy and terms, account deletion, report and block).

What we learned

Splitting one assistant into small, fast triage plus stronger tool-calling and reasoning models makes the whole thing feel quicker, and a configurable model chain made it easy to swap models without touching code.

Built during the submission period

The repository's first commit is 16 October 2026, so everything in it was built for this hackathon window.

What's next

Replanning when money changes, life-event workflows (new job, moving out), a document vault, and data export to CSV, Excel, PDF and Word.

Try it

Built With

Share this project:

Updates

Submission history