# Leapsify

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Minh Tran](https://cerebralvalley.ai/u/leapsify)
- **GitHub:** https://github.com/beenow/spark-voice-trader
- **Demo video:** https://youtu.be/tlo1Zix9KFQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4245

Spark Voice Trader is a hands-free trading assistant that runs on your phone. You talk to it through Meta Ray-Ban glasses or the phone: "How's my Nvidia position, and is it near support?", "Show me next week's Nvidia puts", "Buy two of the second strike". It calls the right market tools, speaks a short natural answer, and drives the app for you: it opens the chart with the support line it just mentioned, the options chain on the strike it named, or a pre-filled order ticket. Orders are never placed by voice: you review the ticket and tap send.

The agent is Meta's open-weight Llama 3.2 1B, fine-tuned with LoRA by distillation from Muse Spark: Spark answered thousands of trader questions with our tools through the Meta Model API, and we kept only the answers with correct tool calls. The model runs a closed loop: it picks the tools, code executes and checks them, and it writes an answer whose numbers are filled in by code, so it never speaks a number it made up. Conversation memory handles follow-ups like "what about its RSI".

Inference runs on the phone: the fine-tuned model, quantized to under a gigabyte, generates every token on a Galaxy Z Fold, with on-device speech recognition and an on-device neural voice. Only the tool calls go over the network, to a service that plays the stock market.

Results on 238 held-out trader questions: the Spark-trained 1B picks the right tools 90.8% of the time (89.9% right arguments), ahead of Muse Glimmer 30B (61.8% base, 77.7% trained on the same Spark traces), and answers in 0.68 s p50 versus 4.2 s. The phone build scores the same on tool choice. The adapter is public on Hugging Face: https://huggingface.co/leapsify/spark-voice-trader-llama-3.2-1b-lora

## More from Meta Global AI Developer Hackathon

- [Parth Patel](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4242)
- [Recovery Room](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4243)
- [Team FarmVision](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4244)
- [SoS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4246)
- [Model Atlas](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4247)
- [Timber](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4248)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4245. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt