Impel was an AI assitant that I worked on for a few months with Impel Intelligence. The idea of Impel was to create an AI that could run completely on-device and interprete what you were doing in real-time. This proccessing was done through a combination of a text-classifciation model that I trained and some good old custom detection code.
Summarizing documents and research papers was one of the most important features of impel. It allows the user to summarize an entire article and then chat with the document to get a better understanding of what is going on. This brought together the benefits of AI being able to summarize documents, as well as explain them in any way needed.
Impel supported full natural language search for every piece of content it ingested. If there is a concept you generally knew, adding some keywords or vagually explaining it would be enough for Impel to find the content you wanted.
This was implemented using a custom Vector Database backed by Faiss along with some clever content tagging and chunking.
Like document summary, video summary allowed you to summarize any Youtube video in seconds. In under 30 seconds I could summarize the entire two hour WWDC keynote and give you a quick rundown of what was released
Impel was able to detect when you were in a meeting and would automatically record and transcribe meetings. This gave you an automatic note taker, no more worrying about taking notes or starting a recording.
Since Impel's search was focused on finding exactly what you needed, it was able to search across a plethora of connected services. Any documents in Google Drive, Notion and more were available to Impel's vector datbaase. This allowed you to have on central serach for all of your knowledge.