What's new in April 2020

New infrastructure, Chrome Plugin, Automated Training Pipeline, Knowledge Base Integration, Template variables and Sentence Splitting.

New infrastructure ☁️

Every month suggestion usage increases and algorithm enhancements require more performance for our AI engines.


With the migration to the new infrastructure our response times are now well below the 100ms industry standard for real time experiences; enabling future growth.


Seamless autoscaling handles the need for more capacity during peaks in the evening and slow hours at night without users noticing.


As an added benefit we are capable of satisfying a host of compliance requirements, leading up to ISO-27001 certification. 🧼

Chrome plugin

Native integration with our plugin


We strive for the most natural way to assist agents in writing their messages. No more looking around for suggestions, they are right where your attention is 👀


With our new plugin we’re able to display our suggestions and autocompletions right where the agent is typing...| This way agents keep focus on their message and, even more important, your customer.


The plugin works natively with any customer contact center platform, from Genesys and Salesforce to Zendesk. This enables seamless integration with your existing workflows in your current contact contact solution.

Automated Training Pipeline

Our GPU trained models are now live in production using TensorFlow serving. 🏎💨


Our pipeline was already updated to train* our models on GPU (instead of CPU) at the beginning of this year. Reducing training times from days to minutes. These GPU trained models are now live in production using TensorFlow serving.

50% increase

With that we have better performing models, with faster inference (for real time suggestions) and we can run new models in A/B, to make sure each new model is an improvement. ⚗️

This also allows us to train on larger data sets. For instance, sentence splitting gave an increase of 50% in training samples.

Automated training

This new feature enables automated training, allowing to train new models more frequently, without intervention. Training frequency can become weekly, or even daily, giving better results. As a next step towards automated training we now use a tailored training database, with continuous integration for the deployment of new models.

Knowledge Base Integration

Most enterprises deploy knowledge bases to structure the way agents determine which action they need to take to resolve customer questions. We now enable a standardised way to integrate this RPA functionality seamlessly within Deepdesk.


We present flows of questions directly as suggestions to the agent, which can be send directly to the customer. Resulting in less writing time for the agent, quick responses to the customer, and faster problem solving. 


The integration is currently in production for use by agents. 🪂 Already a range of ideas on how to improve have come up: like handling the customer’s response automatically and storing the responses directly in CRM systems.

Template variables

Conversations often contain personal information like addresses or names. Deepdesk automatically fills in these values when offering suggestions. Variables that could not be filled are also highlighted now, to make sure the agent fills in the proper details.

Machine Learning: sentence splitting

One of the key metrics of our performance is Coverage: The percentage of the actual chat history that can be written with our suggestions.' Where suggestions exactly match the historic data. We measure this over the chat history of the last 3 months.

Where our algorithm used to look at messages, we now split messages into sentences. In some cases this results in a 40% increase in coverage, more suggestions and better autocomplete matches.


Read more in our January 2020 post 🤓

Header image by Masaaki Komori on Unsplash