Inspiration
Most of us have never been taught how to negotiate. When a salesperson puts us on the spot, it is easy to forget our budget, overlook an important condition, or agree before asking the right questions. Even everyday decisions can be surprisingly difficult: a group of friends can spend more time deciding where to eat than actually eating.
We built outspoken around a simple intuition: knowing what matters to you ahead of time should make decisions easier in the moment. We wanted an assistant that remembers your priorities and offers a discreet cue while you stay present in the conversation.
What it does
outspoken is a wearable conversation assistant built around the Vexo band and a companion web app. It uses conversational context and your confirmed goals to help you evaluate choices.
In negotiation mode, you set priorities such as your budget and required return policy. The assistant interprets offers against those requirements and provides a simple recommendation: proceed, pause, or ask for missing information. The companion app shows the transcript and reasoning, while the band provides an at-a-glance color cue.
For everyday choices, such as picking lunch with friends, it helps compare three options against the preferences and constraints you provide. The goal is to turn a circular discussion into a concrete next step.
The prototype also includes a gesture-triggered voice update flow: wave three times, speak a change to your goals, and review the proposed update before confirming it.
How we built it
We used Rust for the application backend and orchestration, connected to a device bridge for the Vexo band’s microphone, motion sensors, and RGB LED. Deepgram provides speech transcription, and Google Gemini interprets conversational information into structured terms that the application can evaluate against confirmed goals.
The companion web app lets users configure goals, review transcripts and recommendations, and confirm proposed changes. We track sessions, turns, and goal revisions so that recommendations can be associated with the conversation and preferences that produced them.
Our architecture separates interpreting an offer from checking whether it satisfies the user’s requirements. A missing fact should prompt a question rather than become an assumed agreement.
Challenges we ran into
The hardest part was keeping speech, context, and feedback synchronized. Transcription arrives asynchronously, people pause halfway through a thought, and goals can change while a recommendation is being generated. We encountered stale-response and voice-capture issues that showed how much care this coordination requires.
We also had to translate detailed reasoning into a useful signal on hardware with no screen or speaker. Integrating the device bridge, transcription service, backend, and web interface under a hackathon deadline meant continually narrowing our scope toward a clear demonstration.
Accomplishments that we're proud of
We built a prototype spanning wearable input, speech transcription, structured offer interpretation, goal checking, and a companion interface. We are especially proud of making user-confirmed priorities central to the design and requiring confirmation before spoken updates replace active goals.
We also developed two relatable use cases—negotiating a purchase and choosing lunch—to explore how the same idea can support both consequential and everyday decisions.
What we learned
A useful decision assistant needs to recognize uncertainty. “Ask about the return policy” can be more helpful than an immediate yes or no.
We learned that wearable AI depends as much on timing and interaction design as model quality. A recommendation needs to reflect the current conversation, arrive at the right moment, and be understandable without demanding all of the user’s attention.
What's next for outspoken
Next, we want to harden the live hardware workflow, reduce response latency, and improve voice updates across pauses and interruptions. We also want to test the cues with real users to understand when they help and when they distract.
Beyond the prototype, we envision richer group preference handling for shared decisions, more personalized negotiation support, and clearer controls over recording, retained context, and deletion. Our aim is to help people make choices they understand and feel confident owning.
