What's new in March 2021

✨ Personal Collection ✨ New benefit tracking
✨ The first Dutch GPT-2 base model ✨
And of course ⚒️ under the hood.

Personal collection


Agents often have a doc or notepad with frequently used messages. What if that would work with the power of Deepdesk? That's what Personal Collection is.

Add all your favourite messages

You can add as many messages as you like. When you start typing a reply to the customer: they're right there.

AI powered

Of course, when you use them enough, they will be picked up by the model. So that they are recommended just when you need them.

Private beta

Personal collection is currently in use by a limited group of agents. And preliminary results are quite promising! Agents are happy to have a sense of control. And data shows a significant increase in Deepdesk usage. Sign up for access to the public beta.

The first Dutch GPT-2 base model, doubling performance

We’ve picked up the audacious task to train the first Dutch GPT-2 base model from scratch, Since no Dutch base model existed, we had to do it ourselves.

The GPT models we already had in production for our customers, were fine-tuned on an English base model. Of course we expected a Dutch base model to produce even more useful recommendations.

It took some effort to get a large corpus of Dutch text, and to setup the infrastructure and code to run the training. When it eventually succeeded, the training ran for over 12 days. Processing 65GB of Dutch text.

The Dutch base model was then used to train a model for one of our customers, fine-tuned on their chat data. It currently runs in A/B, with the new model outperforming the previous model, resulting in nearly twice as much text in used suggestions.

Proof of value

The key metric we use to indicate how well we do is Assisted Text: the percentage of text that is assisted by Deepdesk, and thus not written by the agent.

Customer KPI

Although it’s an adequate metric to indicate our performance, it does not directly indicate the benefit for our customers. Customers expect Deepdesk to have impact on either Average Handling Time, Chats/hour or any similar metric that expresses agent or contact center performance.

Measuring handling times

We extended our metrics and now also track handling times. So that we can demonstrate the impact Deepdesk has on their own key performance indicator.

🛠 Technical

We are putting a lot of effort into slowly breaking our giant monolith down into smaller parts, to improve the maintainability, and to streamline the CI/CD pipelines. This also allows for faster development and release cycles for the individual services.

By adding Sentry monitoring and alerting to our frontend, we now have a much better awareness and understanding of problems occuring in the browser.

Our main focus was further automation of the onboarding process. We have started using Helm to manage our Kubernetes deployments, and Terraform to provision the infrastructure needed for new customers. These efforts have also made it possible to create strict isolation between tenant infrastructure, further improving the platform security and compliance.

More?

Read more in our winter 2020 post ❄️

Stay up to date

Header image by Magnus Jonasson on Unsplash