NylWave

NylWave: Make your emails more productive with AI.

With NylWave you don't have to copy your email texts to different tools, but use predefined or custom prompts directly in your email client.

Inspiration

With the hype of Generative AI, especially Large Language Models like ChatGPT or Lama2, we see a lot of possibilities to increase productivity. However, we see - and want to challenge - two pain points:

a) Integration: Currently, most users access LLMs by opening another tab or having an external tool opened, needing to copy and paste texts, leading to more context switching than required. b) Privacy: While cloud-hosted LLMs like ChatGPT are easy to use and very popular, it requires the users to send internal information to external providers. For that reason, even some large companies have banned ChatGPT for their employees' usage.

What it does

With NylWave we address both points: We integrate LLMs directly in your email client. Via the Nylas-API, we integrate that not in just one, but all email clients at once. We utilize the deltas-API to detect creations of drafts and edit them if certain commands are detected. Instead of using cloud hosted LLMs, we deploy GPT4All to guarantee privacy, if NylWave is hosted on customer Hardware.

We integrated the following commands:

• # NylWave: add quote: adds an inspiring quote at the end of the email.
• # NylWave: make formal: make the email sound more formal.
• # NylWave: correct errors: run some automated spell & grammatical checking.
• # NylWave: prompt <prompt>: directly access a Large Language Model with a custom prompt.

How we built it

We created a Webservice that allows you to register (and de-register) email accounts. As long as an account is registered, a background worker is listening to draft edits and modifies them whenever a command is detected.

Challenges we ran into

Since we are using a 3 Billion parameter model instead of a 175 Billion model, we noticed some prompts were not working as well as we thought initially. Due to that, we could not integrate fun commands like "catifying" an email (adding cat related puns). However, all formality related prompts were possible.

What we learned and we're proud of

We learned to set up a LLM by ourselves, and while we have a lot of experience with AI, we didn't have experience with models sized of multiple billion parameters.

What's next for NylWave

We consider the following advancements to our idea:

• Organization wide customizable prompt templates: Assuming that NylWave will be deployed for an organization, the commands and prompts can be customized and shared across teams.
• Additional context for prompts: Currently, the LLM has access to the email body and the subject. However, this could be extended by the sender & receiver names, calendar entries, use case-specific APIs, etc. This would lead to the LLMs being able to provide even better E-Mail templates.
• Allowing different (or customized) LLMs: For larger organizations, the LLM can be fine-tuned on the private company data to yield better results.
• More prompt commands: While we prefer giving the user the ability to configure their own prompt commands, a wider range of commands will cover even more of the common use cases.

Built With

• docker
• fastapi
• gpt4all
• postgresql
• python
• sqlalchemy