Inspiration

Most AI companions are built to love you on day one. We wanted one that starts knowing nothing and grows only from what you teach it. When T-Mobile asked for new ways to protect people from scams, we realized a friend kids already trust is the best one to teach them to stop and check.

What it does

Mochi is a 3D sheep who lives in your browser and learns from you. Ask for something or draw it, and he builds it and uses it. You can also pose his arms to teach him a move. Paste a suspicious text and he plays detective: he gives a verdict, says how sure he is, shows the warning signs and the next steps, and hides your personal details first. A scam-spotting game and skits with his friends Mimi and Cocoa teach kids the warning signs.

How we built it

Stage: three.js in the browser. Server: Node.js with SQLite memory on one mini PC. Chat: Gemma 4 26B running locally in LM Studio. Scam reading and prop practice: NVIDIA Nemotron 3 Nano on a Brev GPU. Prop fixing: Gemini, for props that come out wrong. Feelings: a RoBERTa model we fine-tuned ourselves. Voice: Kokoro.

Challenges we ran into

With thinking on, Nemotron took about 200 seconds per prop, so we moved it to scam reading with thinking off, which takes about 2 seconds. Our first generated dataset was flawed and had to be rebuilt. Two checks gave too many false alarms until we added a third vote.

Accomplishments that we're proud of

On 408 messages Mochi had never seen, he caught 97% of scams, and 94% of his alarms were real scams. When he said “pretty sure”, he was right 323 of 327 times. Personal details are hidden before anything leaves the computer.

What we learned

Give each model the job it’s best at. Rules, a small trained model and a large model together beat any one of them. Measure confidence before you show it.

What's next for Tamamochi

A call coach that listens to a call on speaker and warns you live. Checking screenshots of texts. Testing on real labelled messages. A summary page for parents.

Built With

Share this project:

Updates

Submission history