PowerPoint to Email with OpenAI

I was having a conversation with a colleague during his recent visit to the U.S. and he was mentioning efforts to fight back around the constant meetings that were filling up the calendars of many in his team.? Meetings serve different purposes, some of them extremely useful, and some inefficient. Maybe the prime example of this are the sort of informational meetings that many attend without contributing to the discussion. In these cases, aren’t there better ways to handle things?

After our conversation, I was thinking back to the old joke “this meeting could have been an email” and thought, why not make that literal? ??

So, here’s the recipe for this:

1. Take the PowerPoint presentation that someone prepared for the meeting.
2. Yank out the text from each slide (with the pptx library in Python)
3. Ask the openAI API to summarize each slide as a sentence
4. Concatenate those sentences
5. Ask the openAI API to summarize all that information as an email

I’ve tried out a test case (doing a search on “biotech non-confidential slide filetype:pptx” wound up returning this older presentation from ESSA).? And…after processing through the above process, here’s the result:

This powerpoint is about ESSA Pharma, a publicly traded pharmaceutical company that is focused on developing novel therapies to treat prostate and hormone-driven cancers. The company has secured a $35 million ATM facility with Jefferies and is planning additional preclinical studies to fund their 2020 to 2023 program. Their current drug, EPI-7386, is a unique inhibitor of the N-terminal domain of the AR with improved potency and pharmaceutical properties compared to EPI-506. The company is currently conducting a clinical study involving 19 patients with advanced solid tumors to evaluate the efficacy of a monotherapy and combination treatment using the drug EPI-7386. Customer service is available by contacting Peter Virsik, the COO of Essa Pharma.

Not too shabby!?

A few things are suboptimal about this - I’m pretty sure it wouldn’t work for very long presentations without chunking, due to token limits in the openAI calls. It’s also reliant on the text itself in the PowerPoint being interpretable - I’d have to bolt-on additional tools (such as those in Azure) to describe images to capture presentations that are heavy in that direction.? And, of course, this is going through openAI APIs, so you are out of luck using this for any confidential or sensitive information.

If you’re interested in trying it yourself, let me know in the comments and I can throw the code snippet on Github.