In this series we will shed light on the extraordinary community work done by our Rasa Superheroes, take a deep dive into what inspired them, and find out if they have any tips or advice to share with aspiring community builders and conversational AI enthusiasts. For our second instalment, we get to know Souvik Ghosh.
Souvik has an outstanding track record of supporting our community. He sits comfortably in the top 5 for “most replies” and “❤️” reactions of all time on our community forum (only matched by our team!). That’s not all, he is also the only Rasa community member to obtain the Aficionado badge, which is only achieved by visiting the forum for 100 consecutive days! There is no doubt about his dedication to help out our community of open source developers.
Living in Belgium for the past 5 years, he’s currently working as a machine learning engineer for BNP Paribas Fortis. There he manages and builds industrialization pipelines for analytical models on the advanced Data Analytics platform. We spoke to him about his current projects, ethical bot development, and transcending language barriers to make AI assistants that work for everyone.
Hi Souvik! What does a typical day look like for you?
An average working day includes tons of coffee and managing a fairly large infrastructure that supports our data science teams and takes their models to production. Outside of average work hours, I am involved with early stage startups trying to jump on the European scene, collaborating with them on new trends, technologies, and their potential application to solve the problems these companies are after. Having moved a bit beyond NLP, recently I have been focusing a lot of my energy in DevOps for ML products, or MLOps. In particular, continuous integration/continuous deployment, as it's a key area for machine learning today, more so than traditional software development.
You clearly have a fantastic background in computer engineering, but what inspired you to get into the conversational AI?
I do come from a computer engineering background, moving towards Information Systems. I quickly realized that the so-called Industry 4.0 will have a huge impact in the financial industry. There were internal calls for developers to join new innovative projects, which was my cue to jump on board!
Do you remember how you found out about Rasa, and how you started contributing?
Dealing with a lot of very sensitive data, we have certain constraints with the type of solutions we can go with. Open-source solutions were the upcoming trend back in 2017, and a quick GitHub scan led me to some very interesting projects.
Rasa NLU (now merged with Rasa Core and known as Rasa Open Source) stood out, due to its flexibility to modify the pipeline and this catered perfectly to our needs. Given that we live in a country with 3 official languages (Dutch, French and German), it was imperative that we found a solution that fulfilled this requirement.
As Rasa was growing, so were we, trying to learn the tradeoffs and advantages of building out our solution based on it. I felt that most of what we do within the company would be useful to share with the community as well - that’s when I started writing tutorials online trying to help the community with questions or problems we have faced ourselves too.
In 2018, Souvik has written tutorials on:
- Building a Rasa Chatbot in Bengali using Supervised Word vectors from scratch
- Build a Rasa NLU Chatbot with spaCy and FastText
- Building a multi-lingual chatbot using Rasa and chatfuel
All of which provided guidance on building bots that improve accessibility to localized sources on the digital market.
Later, he dives into the fundamentals of Rasa Core, in “Contextual Conversational Engine— The Rasa Core Approach” Part 1 and Part 2.
You have been a huge support to the Rasa community since early 2018. What inspires you to do this?
Contributing to such a project means that I am somehow involved in revolutionizing the internet as we currently know and interact with it!
I believe that communication systems such as chatbots are the key that will allow more and more users to be able to interact and access a plethora of services from across the world, from breaking down language barriers to improving accessibility to knowledge and information.
It’s exciting to be part of an open-source project that is having a great impact on the way larger populations are interacting with internet services, making interacting with these services much easier and also fun!
What can you tell us about the projects you have been involved in?
I am unable to give full disclosure on all of my projects in respect of confidentiality agreements. However, I am really excited about a side project I am currently involved in with a startup who are doing some great work in validating the concept of Consent Management and Transparency. Effectively, how analytics and AI in general can benefit us as individuals and why GDPR will become the highlight of Europe’s journey towards AI in the next decade.
Speaking of the new privacy laws, what were the most challenging aspects of navigating GDPR for ML?
Procuring training data that is close to real conversations!
Since GDPR became a thing, privacy law on text mining has been subject to consent, and back in 2018 when it was newly introduced, there was not a lot of clarity on what we can and cannot use. Hence, we decided not to use real customer conversations to train the bot, and this along with the size of our organization resulted in us getting creative. It was possible to sign up a few employee volunteers, and with their consent, we could get more realistic training data. We are still navigating this even today.
Another interesting but unrelated challenge was CI/CD for ML models. Because we were continuously training models each day, redeploying them meant that it was quite a challenge to be able to continuously improve the bot. We incorporated the advantages of distributed systems like Docker, Kubernetes and some good ol’ Jenkins to build out our own CI/CD pipelines to deploy models as we trained them.
There are still so many challenges and learnings to be made in the conversational AI domain– do you have a memorable piece of advice to share from your early days of building chatbots?
One of the best pieces of advice I’ve heard, is that chatbots are part of an automation ecosystem, and that we should really consider augmented intelligence over artificial intelligence to enhance the capabilities of customer service agents. Another learning is that often we should focus on solving a particular problem rather than looking at it as a solution to every problem.
I still provide this same input to other early stage startups about focusing on bringing customer value while building their product. We have millions of customers who trust us to provide them with the best of aftercare. After all, being in retail means that customers are our top priority, and at the heart, we are a human centric company.
How do you think technologies like Rasa can change the world?
One of THE biggest factors that really differentiates Rasa is the inclusiveness of other non-traditional languages of the Internet. The fact that the pipeline is so flexible, allowing developers to build out bots in literally any language, is unprecedented. Larger cloud frameworks, back in the day, only supported a handful of European language. Rasa allowed me to build bots in Swahili, French, Dutch and even Hindi and Bengali. Such flexibility allows local services in several countries to be more accessible and enhances the user’s ability to interact with modern internet systems in a much more efficient manner.
If you move across the large pond towards countries in Asia or Africa, there has been a tremendous growth in the mobile-first economy, and I think that chatbots will play a crucial role there. Building out systems in native languages will help these countries grow and be more inclusive in the process. Even though English is still the most dominant language of the internet, it can hinder growth in countries where English is not very common. This is where NLP, and especially Rasa, comes in.
Thinking about where we are right now and looking to the future, what are some of the biggest trends you see in conversational AI?
NLP has really grown over the past couple of years, and more and more focus has been put into productizing these new frameworks for real-life situations, which is exciting. Models like Transformers, BERT, and others are now being miniaturized to allow them to be put into production. I believe this really opens up a lot of doors for Natural Language Processing as a domain in the future.
We had the pleasure of getting your take on the world conversational AI, to wrap this up, what do you do when you’re not building and managing those pipelines for analytical models?
Personally, I'm a big fan of cooking. Having lived with people from many parts of the world, it was always through food that I found my true connection. One of my most favourite meals, which I learned from a roommate of mine, was “Buffalo Wings”. Quick and easy - mix your wings with some flour, garlic powder, cayenne and salt, pop them in the oven for 20 min each side, toss them up in Frank’s Buffalo Hot sauce and Voilà, delicious party snack :D (I am not paid by anyone to endorse this!).
Join for a spotlight, stay for the spotlight, come away with a new meal recipe... another great reason to follow Souvik!