RedInk

• 

  Thumbnail.

Inspiration

I was inspired by the current global events and how many many influential people and global leaders are stating claims that could cause a lot of damage. The core of this are live speeches, debates and interview, where false claims spread.

What it does

RedInk captures browser tab audio, transcribes it live, and uses a date-aware LLM to instantly display fact-checked claims as vintage ink stamps in a newspaper-styled sidebar.

How we built it

I built a lightweight Manifest V3 extension using vanilla JS, piping captured audio to Deepgram via WebSockets and evaluating claims using small models on Featherless API gateway.

Challenges we ran into

I had to overcome browser WebSocket header limitations for Deepgram authentication and implement speech buffering to ensure the LLM evaluated full sentences rather than fragments. Also finding free APIs and models was also a hurdle.

Accomplishments that we're proud of

I am proud of achieving sub-3-second end-to-end latency with a beautiful, zero-dependency broadsheet design isolated inside a Shadow DOM.

What we learned

I learned how to securely handle stream capture using googles audio API and constrain smaller models to reliably return structured JSON.

What's next for RedInk

Next I want to use stronger models that are more accurate and could reflect real world events better. Also using a multi agent architecture could also improve speed and reliability.

Built With

• chrome
• chrome-offscreen-document
• chrome-tabcapture-api
• css
• deepgram-api
• featherless-api
• html
• javascript
• qwen-llm
• websockets