“Create notes with your voice. Like magic.”
Conjure.Page is the productised form of how I’ve worked for the last several years. I’ve always been a talker, and since the inception of Whisper voice memos have almost completely replaced my taking of notes. My typical flow for a meeting or brainstorm is this:
- Record audio
- Transcribe with Whisper
- Feed into LLM to summarise / refine
- Use refined notes within multiple LLM conversations to further brainstorm/discuss points
Conjure supercharges this by jumping straight from voice memos to finished markdown documents that include not just text but diagrams and images too. It’s currently in a state where I can generate polished documents with text and some diagrams with images TBC (😭). The main requirement for this product was that the rendered document was holistic and non-linear, meaning that the rendered document was a representation of extracted data up to time T, rather than a traditional transcription. This would mean that one statement may completely change the entire document, which is what I wanted as my voice memos are usually me working through a problem and adapting/evolving my thinking over the course of the memo.
The main technical breakthrough on this project has been in creating an encoder/decoder architecture for document generation. We need to go from a very unstructured body of text into a structured document that has cohesive composition and narrative. This involved the use of knowledge graphs, XML and multi-agents.
I’ve put the project on pause for now but will likely release sometime in late 2024.
