Hold That Thought

Imagine you're talking with a friend, and you keep having to search stuff up; well, hold that thought, because Hold That Thought saves you from the hassle by showing you info about your conversation.

Comment

Inspiration

One of our friends was talking to his cousin about cars, and he hated the hassle of having to search up facts about said cars all the time; why not have an AI listening in to handle this for you?

What it does

Hold That Thought is a real-time conversational assistant delivered as an iOS widget/Live Activity.

  • Listens to ongoing conversations using on-device or low-latency transcription
  • Extracts key topics, entities, and factual claims in near-real time
  • Uses an LLM to infer what information might be useful right now
  • Surfaces concise facts, clarifications, definitions, or follow-up prompts
  • Runs as a Live Activity, so insights appear without opening an app. The result is a subtle, glanceable stream of context that helps users stay informed during conversations without breaking the flow to open a search engine.

How we built it

The app is built with React Native + Expo, using native iOS capabilities for microphone access and background execution. Audio is streamed in real time to Deepgram Flux for low-latency, incremental speech-to-text and conversational recognition. Transcripts are processed as a rolling window and fed into an LLM-based inference pipeline that extracts topics, entities, and implicit questions using structured outputs instead of long-form generation. Insights are delivered through iOS Live Activities, as well as the main app screen, allowing users to see context without opening the app if need be.

Challenges we ran into

  • Minimizing latency: The app is API-call heavy; we spent a lot of time minimizing latency for a smooth, near-real-time experience.
  • iOS struggles: iOS was really finicky to get working, and puts a lot of constraints on developers, especially for Live Activities.
  • Keeping the assistant silent: The assistant was often too eager; maintaining quality over quantity was a challenge, but we found a healthy medium.

What's next for Hold That Thought

Releasing on the App Store, of course :) (JK) We tried to get a chat history feature in, and almost did, but couldn't in time. QoL features such as those would be our next focus.

Built With

Share this project:

Updates