GOAT++

Floripa merchant using the app

Inspiration

From my personal experience, I see many people, including me, struggle to confidently use Bitcoin software and features. GoBTC tries to solve it with human expert volunteers, working for Bitcoin adoption, but it doesn't scale. But we can scale with AI. After some research, we saw that most merchants and users GoBTC helps are beginners and are in need for this kind of tool. The inspiration for the name is Fran Finney.

What it does

Unlike a general chatbot like ChatGPT, Fran can see your screen and talk to you, so you get an as close personal experience as possible. You can ask her questions, she can guide you through difficult UIs and suggest commands to type without the need to be changing screens all the time.

Fran have lots of tools to be able to understand and help most with important issues someone has.

How we built it

We used Gemini 2.0 real-time api for a multimodal agentic LLM starting from a react/typescript example. We built some specific tools for specialized agent function calling and elaborate prompt engineering for objective steering.

Challenges we ran into

Using Gemini is not ideal for bitcoiners because it is closed source and all information goes through Google, but it is enough for a 36h hackathon PoC. Also closed source models can change over time and become "dumber" and unpredictable, leading to reducing the accuracy of the responses.

Open source models would be challenging to connect real time the SpeechToText and TextToSpeech modalities.

Gemini Portuguese TTS is also definitely not great which will force us to recreate Fran from scratch using another LLM and a specific good TTS PTBR model.

Lastly, we don't code in typescript, so it was particularly challenging to create this since the only starting example was in TS.

Accomplishments that we're proud of

We already had a first user, a Floripa Merchant, trying our product. There are 300 others like him onboarded by GoBTC who would greatly benefit from using it.

What we learned

Even hardcore technical Bitcoiners have it tough from time to time, needing to reach to the depths of Bitcoin stack overflow for answers to their Bitcoin problems. This means that despite Fran being built for helping Bitcoin adoption from merchants and newcomer users, adding more complex and technical functionality will help experienced users as well, as reported from our field research.

What's next for Fran

Bitcoiners not always feel safe sharing their data with google. A logical next step is to create use local open source models, so the user can be 100% safe no one is going to use their data, and also be able to use Fran offline.

Also it will be very useful for GoBTC to quickly adding merchants by voice to BTCMaps when integrating with its API.

Also integration with Bitcoin knowledge base, making Fran the most Bitcoin savvy model out there.

NWC for tipping the best answers.

Also generating transcripts for easy copy/paste commands.