Inboxity

Inspiration

Perplexity deep research is powerful tool for productivity, yet most people don't use it at work. I realised that in order to solve this, we need to meet professionals where they are: their email inbox!

What it does

Inboxity is an AI research intern that lives in your email inbox. You simply send them an email with a research request, and they will respond within a few minutes with a deeply researched, fully cited, professionally formatted pdf report.

If you wish, you can then forward the pdf report to others to show them what perplexity can do. They are also welcome to email your intern for reports of their own.

How we built it

I used a pipeline of technologies to create Inboxity.

1. Whenever an email is sent to my intern's email address, A1Base will notify me by sending a webhook to my express server
2. We then forward the request to Perplexity Deep Research API (sonar-deep-research)
3. After receiving the deep research response with citations, we send it to the sonar-pro model. This model will format the report as HTML and also search for any relevant images.
4. We then use puppeteer to convert the HTML into a PDF.
5. We upload the PDF into a supabase bucket and get the public link.
6. We use A1Base to send the PDF report as a reply to the original email.

Challenges we ran into

Initially, I was trying to convert the deep research markdown reports into pdfs deterministically by developing a mapping from markdown to pdf elements. This did not look good as the formatting was very plain and never adapted to the content. After struggling for a while I realised I could add a second step to the AI pipeline where I use sonar-pro to convert the markdown into stylised HTML. This solved the formatting issues entirely as sonar could generate nicely formatted HTML, and had the additional benefit of allowing me to enrich my reports with relevant images that sonar found on the web.

Accomplishments that we're proud of

The deep research to pdf pipeline is a pretty cool solution, in my opinion. Ended up being a really cool way to convert plain markdown into a nicely formatted PDF report with relevant images.

What we learned

I learned the AI engineering technique of using proxy formats. i.e. When trying to convert from A->B with an LLM, it can sometimes be easier/better to convert from A->C, and then from C->B. And often using an LLM for one of these conversion steps can give you a level of control that deterministic code struggles with.

Regarding perplexity, I also learned:

• How to use the deep research API (very simple, nice!)
• The technique of using sonar-pro to both transform AND enrich data (in my case, I converted from Markdown to HTML, and enriched it with relevant images found on the web).

What's next for Inboxity

Deep research over email has so much potential. I have barely scratched the surface.

Some ideas I have brainstormed with friends include:

1. Better interactive experience Simple email comms are a good start, but you should be able to chat to your intern over Slack. Not only that, but the intern should ask follow up questions to clarify, and you should be able to request edits or upload internal documents as context.

2. Generating deep research reports on a schedule For example, you could ask the intern to "email me a monthly report on any new products that my competitors have launched". So much potential here for scheduled report generation.

3. Custom email newsletters with Perplexity Deep Research! Imagine asking your intern to send you a weekly email newsletter on any topic of your choice, no matter how niche! Just describe your ideal newsletter to the AI intern and set a schedule.

Built With

• a1base
• express.js
• perplexity
• supabase
• typescript