Why we removed Machine Learning from our pitch-deck

Why we removed Machine Learning from our pitch-deck

This story starts with a tweet regarding the use of Machine Learning and Artificial Intelligence in our pitch-deck, presentations and casual chit chats on what we are building at Feedgrip.

We indeed ❤️ Machine Learning and Data Science. In fact, 40% of our team are Data Scientists (this means 2 of 5) with a great passion for the field. To be more precise, this whole thing sprang from our eagerness to automate tasks that Product Managers perform daily regarding User Feedback management and the path from user’s voice to product development.

Our first take in tackling the problem of user feedback management was to promise our potential customers automatic clustering and insights from all feedbacks. Given enough volume of text-based feedback, state of the art algorithms can produce magic outcomes.


We experimented with curated and synthetic datasets in the scale of hundreds of thousands of feedback documents. We were able to mine patterns, associations, and trending topics. We have code in place that does sentiment analysis of user feedback and produces sentiment timeline graphs. We are able to tell whether a user sends a bug report or feature request. Our focus was information retrieval and topic modelling from short text. However, none of this found any application in the stage we are currently in.

After numerous talks we had with Product Managers (described in detail here) we realised one thing:

Users do not care about your technology

Specifically, users only know their problems. Users only care about their problems. They do not know the solution and are too busy to think of one. They couldn’t care less about the details of the specific solution we came up with.

After several months of experimentation with exotic technology, be it Deep Learning, Convolutional Neural Networks, Recurrent ones, Embeddings or multi-dimensional representations of words and entities, we realised that technology should only be a tool in our belt.

Having technology as a main pitch argument, as sexy as it may have been, derailed the discussion from the actual problem we have been trying to solve. It mostly shows our aptness in studying, implementing or even extending state of the art papers.

We don’t care about hype-driven investors

We know that there are investors who only invest in A.I. and Deep Learning. We don’t care about this type of investors. They will not help us and they will go away when another exotic technology arises. We only want to collaborate with people who know why what we are doing here is important and know the Customer Support and Product Development domains very well.

In fact, we decided to remove every single bit of Machine Learning from our MVP. In order to deeply understand the problem and our users, we prefer to tag, classify and organise stuff manually instead of having some algorithm do it for us.

We even prefer having our users invest some more time in our platform (we know this may be regarded as a conversion killer — again, we don’t care) and manually do some stuff, hoping that they will provide better feedback and request missing features. This way we will be able to actually talk with our users, have them tell us that this task and that task are time-consuming and then optimise these specific tasks, avoiding premature optimisation.

Another reason for taking a slight turn away from advertising our Data Science at the moment is the fact that we should start treating Data Science as a commodity. We could not imagine telling our potential users or investors that we use Databases or Ruby on Rails or the Internet. The same applies to Data Science. It’s here to help solve some problems (not all of them) and it’s not an end in itself.

From Machine Learning to Problem Solving

Supervised, unsupervised, active, reinforcement, machine and deep learning should just be configurations of a component. Other components include data fusion, MVC, decorators, serializers, back-up strategies, mailers, etc. All these components together should constitute a black box, namely Feedgrip, whose sole purpose is to make Product Managers’ lives easier and more productive.

Machine Learning will certainly be at the heart of Feedgrip, after all it’s still our passion. Just a less important passion than solving the real problem right now.

Thanks for reading my article. As always, we are Feedgrip. Feel free to drop your email if you are interested in getting early access or beta testing our solution. We will not spam. If you want to learn more about how we work, feel free to drop me a line.

Dimitris Kotsakos
Authored by Dimitris Kotsakos
Software engineer and avid reader.