Last year we invested in Lengoo — founded in 2014, Lengoo leverages its enterprise customers’ proprietary past translation data and a human-in-the-loop approach to train custom, multilingual Large Language Models. Within their 1st skill “translation”, Lengoo’s technology constantly outperforms the models of large tech giants across 50+ languages saving 7-digit figures in professional translation budgets.

We are thrilled to be sitting down with co-founder and CEO, Christopher to shine a spotlight on Lengoo and share his advice for aspiring founders.

Share the story of Lengoo, how did it come into fruition?

I think multiple factors came into play here. The preface to the story is that I was working as a translator, and then later as a localization manager for Accenture — that’s how I got into this translation space. Around the same time, I got the great opportunity to study data science in the US, which was a hugely insightful opportunity as they were a lot further along than we were in Germany at the time. One of the key things I took away from there was that language data would somehow drive this AI revolution. At the time, this was not a popular opinion, but I think now with Chat GPT, everyone has come around to agree on this a little. It was such a big hypothesis at the time as back in 2014, the only commercially relevant applications in the machine learning space were related to image.

Our idea was always very simple: to teach a machine human language.

We teach kids to first listen to their parents and then to speak. Then we send them with this skill set to school to learn all sorts of other skills, and we wanted to replicate this in the machine learning space for the enterprise market.

So our first task was teaching the machine corporate language — in our case for the enterprise segment — and then later we can teach the machine all sorts of other skills to replicate what human translators can achieve.

The next big question was, “ok, we need language data for this — where do we get it?”. I had experience in the language translation space, so we started there originally, with an approach to collect language data from customers through a language service. It took quite some time to figure that out, and we are still learning every day, but we now understand how to serve enterprises with professional translations.

In 2018 we started our research on machine translation. We wanted to start with a skill where we could immediately evaluate the quality of these machines, so we built our technology with the help of EU funding over two years. Then, in 2020 we onboarded our first customer with an attractive value proposition. With all these AI applications you need to give customers a short-term ROI, so we said, look, we’re going to do exactly the same thing your current provider does, but we can do it with the support of a machine, and we can do this three times faster at half the price. This worked well. In the beginning, we had a lot of SME clients and over time we transitioned to enterprise only. We’re currently servicing around 40–50 enterprise clients with this same value proposition.

Towards the second half of last year, we began to understand that we could delve inter other skills beyond translation faster than we had initially anticipated, it was becoming clear that the hype we’re experiencing now was on the horizon. What was interesting for us is that every machine learning area seemed to be moving towards a deep neural network architecture called “transformer models”, something we’d been using since the very beginning for our machine translation. So, all our tech stack from data cleaning, training models, deploying models, and continuously fine-tuning them with human feedback from our professional linguists, was optimised for this technology.

With nearly every field in machine learning from image, video, audio, and text moving towards transformer models, we realised if we’re already training for our customers, these customer-specific, large language models for translation — and they understand how a company likes to express itself in a certain context and can produce much better translations than any other solution out there — we should use this now to expand it to other skills and that’s what we’re currently working on.

We’re also guiding our customers through the current hype to show which use cases are sustainably providing value to enterprise customers. While generating a pink elephant picture is fun, there’s not always much value behind it.

A lot of serendipity and a big bet in the early days that language data would fuel the AI revolution has helped make Lengoo what it is today. We’ve seen a lot of indications by now that this current buzz around AI and Large Language Models is a bit of an “iPhone moment” as opposed to an empty hype, and we’re very excited to be at the centre of this.

What has been the hardest lesson learned so far?

There have been several hard lessons so far, the earlier stages required a lot of learning and some pivoting for us to properly address the market. One example is that when we first started, we worked with students to generate language data. We needed experts, so we figured Ph.D. students writing papers on specific topics would be the best translators, but, we missed a key component — which is that students have exams or semester breaks at the same times, during which they wouldn’t accept any jobs which didn’t make our customers very happy. So, we quickly switched over to professional translators to prevent gaps in service.

Another thing we thought would be logical, was to create a matchmaking system — we have all these experts with certain linguistic and domain expertise, and we have the problem of clients needing a certain translation for a specific handbook, why not let them pick who helps? While this is great for the SME market, by the time you get to an enterprise level we realised that selecting your own translator for every job is additional administrative work that enterprises don’t want. Our largest client now sends us around 400 jobs a day, such a marketplace model isn’t suited.

It’s important to take some time to find the right solution to one problem.

What has been the biggest surprise?

Until today, it is astonishing to me that we can comfortably say that we have the best machine translation solution out there. We’re a small team, we have 30 in our R&D department, with 5–7 people actually doing research, and yet it’s been possible to build a better solution than any of the big tech companies.

If you focus on a very small problem, and you focus everyone on that, it’s possible to out-compete large tech giants, with much more money and more expensive talent.

What advice would you offer other growth companies about financing their businesses?

The first thing is you need to be aware of all the instruments that are available to you. We started Lengoo during our studies, we had no clue about fundraising. In the beginning, we joined TechStars, and they were able to introduce us to the fundraising side of things and show us how the dynamics work. Personally, I found this very useful, but it steered us to focus on equity funding. We had our seed round in 2016 and it wasn’t until the end of 2018 we got our first public grant. I hadn’t been aware of how much money is out there from all these public or government grants. I’d also never heard of venture debt until I started working with Atempo. It didn’t seem to be a widely discussed instrument, particularly across Europe — which is a shame because, for founders, it’s brilliant.

If you’re not completely there yet to raise a new round, venture debt can help you bridge a little until you can increase your valuation ready to raise the next round. And from a dilution perspective, it’s one of the best things founders can do.

If you consider that you’re trying to build a billion-dollar company, you should look closely at your equity and do whatever you can do to prolong the runway. This is something I think should be considered at every funding round, and I wish I had known about it earlier.

What is Lengoo’s biggest achievement so far?

We’re not that young, we started in 2014 and finished our studies in 2016, so we’ve been working on Lengoo now for 7/8 years, but still, we are serving some of the largest companies worldwide with our technology, and this is really cool! I could not have envisioned that really big companies are going to use our language AI and be buying into our vision when I started.

I think that the biggest achievement is that when we began, we foresaw language driving the AI revolution. That if you focus on one skill, such as translation, you will not be limited to offer other skills on top of these models.

When we built our first pitch deck for a seed round in 2016, we outlined our plan to show how we would start with translation, automate that, and then move on to other skills, and the plan aimed for this to happen in 2023. We’re currently looking at introducing other skills! The launch of ChatGPT was a phenomenal validation of everything we had been working on. Foreseeing where the machine learning market is going has been a big achievement, and I have to give credit to our CTO, Ahmad Taie for that.

Any tips for aspiring founders?

Being a founder is not as romantic as it sounds, you have to be up for it. It’s your job to deal with all the things no one else wants to deal with — and that’s fine. My co-managing director says I have this ability, that no matter what happens the day before, the next day I am completely reset and ready to do it again, and I think this mentality is important to avoid going crazy in this profession.

Another thing is that you cannot underestimate the impact of executive hires, both in good and bad ways. Hiring someone who isn’t right for the role can lead to losing a year worth of work. Conversely, if you hire a really great talent, it can pull up the entire organisation in terms of maturity and professionalism.

Finally, as I’m focused on product,

I think staying customer centric is key. It’s very easy to fall into a trap, especially when you’re fundraising, of focusing on what you want to achieve, rather than what the customer wants to achieve.

That’s why we made it mandatory that every employee has to have a contact point with a customer at least once a month. I think this is crucial, you need to know who you are serving and what their problems are. This is how you build a great product.