Inspiration

I often think of missing points that I need to cover or want to check whether someone's opinion is valid during meetings so that I can guide the discussion in a better direction. However, it is difficult to do this while speaking and listening during meetings. Therefore, I created an app to help find missing discussion points and check web sources to see whether someone's claim is correct.

What it does

sidethink can receive real-time voice or text during a meeting. It uses speech-to-text with Apple's local STT or the Soniox API. Once it gets text, it converts the text into claims, which are clear and organized opinions based on the actual STT text. Finally, insight cards are generated from the claims to find missing discussion points, fact-check claims, or support creative thinking. Cards are sorted by importance so that users can quickly refer to them during a live meeting.

How we built it

I built it mostly using Codex 5.6 Sol Extra High. I always use a laptop to attend online meetings, so I built it as a macOS app for now. This app has two modes. One is BYOK, where you can use your own API keys to have more control over your data. I also built a backend server with FastAPI to provide on-demand use and more useful reasoning.

Challenges we ran into

The core value of this app is whether we can offer meaningful enough insights for users to adopt it. Therefore, I struggled to create test datasets that contain actual transcription text and meaningful insights that Codex and I created by checking the text. I tuned the prompts and workflows again and again to generate valuable insights during real-time meetings.

Accomplishments that we're proud of

I turned my idea into a beta version of the app. All the code was generated by Codex, and I never looked at the actual code. I just refined the development workflow and defined requirements for Codex to make it write more stable, higher-quality code.

What we learned

Codex is a great buddy for me when building something, but it has several weak points, even with GPT-5.6 Sol Extra High. Codex tends to focus on micro-scoped issues or use workarounds, so I need to direct Codex to look at essential problems and think of permanent solutions. We vibe coders need to understand the characteristics of coding agents to find an ideal workflow for building stable, high-quality applications.

What's next for sidethink

Once I have prepared a TestFlight build of this app, I will run a closed beta test internally to collect feedback and actual data. Although I have created several golden test datasets, I don't think they are enough to offer consistently meaningful insights. Therefore, I want to collect more data to enhance our reasoning logic.

Built With

Share this project:

Updates