Menu
Sign In Pricing Add Podcast

Heiki Riesenkampf

Appearances

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1021.996

Is it 300? When we initially got started, we had a very, let's say, flaky infra in terms of how we scale different streams, how we patch things together. I think I'm proudest of just very sober engineering decisions that led to much more stable infrastructure, that led to quite a lot of improved latency just from

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1052.37

optimizing using kind of 80 20 analysis and so trying to reduce networking time trying to reduce the the total number of api hops that we need to do between the api providers and our servers i think i've applied 80 20 very well and gotten a lot of bang for and not a not too much of engineering effort that's probably my

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1082.211

That's probably what I'm proudest of, not over-engineering the early version of the product.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1097.023

We had this bug where the longer the call got, the longer the delay in the original speech and translated speech was. We could not figure out where that came from. It was a very strange bug. I initially thought there was a memory leak, but then we're using TypeScript as our stack, so I was a little bit surprised if it had been something that I would have had control over.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1124.757

And I threw a lot of weird, different things against the wall to try to chase it down. I knew that there was a feature that I added very early on, which was saving the raw audio from both calls into an audio buffer and then saving the file ourselves at the end of the call. And I never thought that such a... Simple functionality could be the cause of this delay.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1155.957

But then after a month of debugging and just like scratching my head, I just removed two lines of code to build the audio buffer up during the call. And all of a sudden the delay was gone. And I was like, Damn, that was literally like should have been the first candidate to like chase down.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1174.364

I personally feel like those things are much easier to find if you describe your problem and bug to an engineer that has nothing to do with the code. So basically rubber ducking the bug with another engineer. That's probably one of the costliest mistakes in terms of customer questions and dissatisfaction in terms of not amazing demos to date.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1200.391

And that could have been totally avoided had I just approached the bug with someone with a fresh look at the problem. So yeah, collaborate more. That's maybe the ethos of the story.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1226.324

So right now, I am heads down heavy building mode. So we are trying to get our own model out of the door and replace the existing setup while at the same time trying to improve our own infrastructure further. And so now we're on a stage where in addition to building on a daily basis, we are also trying to hire and expand the team here in New York.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1259.324

Our office is located in Soho and we have an in-office culture. And so right now I'm spending quite a bit of my time on interviewing early hires. We are hopefully about to close our technical lead position and I'm interviewing a few machine learning researchers to start work on our own models.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1282.768

It's really a combination of building and then hiring to just have a best in class product that would basically, let's say, never fail to wow the customer away once we get to that stage of the company.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1295.451

And so I think we're going to be heads down building for the next three to four months and then likely raise the next round in order to start getting even more serious about the research and the training our own models part. Anyone building something early stage, I strongly believe that even the early versions of your product could wow the customer.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1321.457

And if people ask me, why did I go into machine learning when I was able to do research in any part of computer science? I personally always felt good. The machine learning was the one part of computer science where if you showed someone what state of the art was able to do, an average person that didn't know how it worked, they would always feel like it's magical.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1347.182

I did research on some like early image generation algorithms using gallons. And any time I would show what I was able to generate back in like 2017 using a neural network, people were blown away. I feel like that magic that machine learning has, it's still there. And that's the reason why I love that space so much. That's the reason why I've stayed in that space so much.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1376.211

That's the same principle that I applied to the products that I built. I really want them to blow people away because that's the only way how to stay apart from the competition and in the sea of boring SaaS apps.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1401.929

The first person that comes to my mind is 100% my mom. My mom has always been one of the biggest inspirations in my life. I always look up to her because I just feel like she has always done a great job of being very high EQ and always understanding what makes people tick. But also at the same time, just being a very kind and hardworking person. That's definitely my idol number one.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1428.942

I think like on a second spot, I would probably just like generally name like a few very close friends of mine that are all in similar spaces like I am building tech companies. And I feel like I have this group of four, five friends that I can bounce any problem off in my life, be it a personal problem, be that a relationship problem, be that a health problem, be that a work problem.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1458.458

They will always take the time to help me out and vice versa. I feel like those people have been with me for years, and I'm sure that we're going to be there for each other for many years to come. And so I would say that small group of friends is definitely my inspiration and support rock number two in my life.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1495.053

The same thing that I mentioned before. I want to see why they care about the problem. And I want to see how either the product blowing me away or their enthusiasm about the problem and how they're solving it blowing me away. We live in an extremely noisy world. And so every entrepreneur out there needs to somehow stand out from the sea of mediocrity.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1526.798

And so I think really the two ways how to do that is either build something amazing or Or just be so excited about what you are about to build to make people excited to join your journey. Be that as a fan, be that as a customer, be that as an employee, be that as an investor. But basically, just blow people away with the first interaction first.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1552.617

And make the world come along with your vision and mission.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

1565.951

Yeah, thank you so much for that, Noah.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

233.677

CommonBase today is a universal speech translator. We help people speak any language in real time and help people connect with each other across language barrier and also cultural barriers. The company was started last year. After having spent some time working on the initial product, I realized I didn't want that to be my life's work.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

261.556

That's after having moved to the US, I realized that my unique lens, my unique personal experience, led me to discover the language problem, especially being here as an immigrant and realize that, hey, this is something that I have personally felt throughout my life needing to adjust to different languages and cultures. And this is something that I want to see as my life's work.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

29.136

People ask me, why did I go into machine learning when I was able to do research in any part of computer science? I personally always felt that machine learning was the one part of computer science where if you showed someone what state of the art was able to do, an average person that didn't know how it worked, they would always feel like it's magical.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

292.058

And so that's the story of how I got to what I'm currently working on.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

320.138

The very first version of CommonBase was built on top of third party APIs. If you think of the problem of speech translation, you have different ways of solving the problem, but at a high level, you need to take the existing speech, you need to transcribe it into text, you need to take the text, translate the text into another language, and turn the translated text into speech again.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

349.461

So that is the initial pipeline of how we built the first version of the product. And the different tools that we relied upon was Twilio for connecting phone calls. And we ended up going with DeepRAM for the speech transcription. Now we're using Llama run on Grok for the translation, and we're using 11 Labs for the speech generation.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

373.811

So that was the stack for the initial MVP that was just to prove the point that real-time speech translation at a very simple level is possible in 2024.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

413.465

I built the product... In a developer environment that I had never used before, I used Replit in order to build the very first version. And Replit was always known as a kind of education tool. What I liked about Replit, how easy it was to invite other people to basically check out your code and basically help you reason about the code. And so that was an interesting tool.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

437.677

I'm pretty sure that we will grow out of Replit. I just want it to be as scrappy as possible and as collaborative as possible. And so that was an interesting choice that I do not regret making. And probably on the photo stack part, I decided to use Twilio instead of basically building a web-based application because a regular phone call is such a standard format for...

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

465.88

Initiating an audio call, I would say looking back at it, I probably would have preferred to build it on top of like two different browser sessions instead of complicating things with bringing the phone connection into it.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

53.616

I feel like that magic that machine learning has, it's still there. That's the reason why I love that space so much. That's the reason why I've stayed in that space so much. That's the same principle that I apply to the products that I build. My name is Hege Riesenkamp, and I'm the founder and CEO of CommonBase.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

601.25

Our product is very different from your average SaaS application. We will or are already becoming a research heavy company. I told you that the first prototype was built on top of third party services. We're now in the process of fine tuning our own machine learning models to do the same pipeline end to end.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

628.317

The next iteration of that would be to actually build our own end-to-end model that we will definitely base on some of the existing research papers that do end-to-end speech-to-speech translation. But all of that is a very active research area. In a way, I know which step we're going to take in terms of increasing the risk and complexity of the machine learning pipeline step by step.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

658.142

But it would be very hard for me to predict how long it would take us to prove each step of the way. And regarding the APIs or how our customers are going to connect to the core piece, which is the speech-to-speech translation, I feel like that part will likely stay relatively stable. And most of the efforts are going to go into just building the best in class speech translator.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

699.841

This is the third venture-backed company that I'm now building, so I feel like I have plenty of battle scars and tiring mistakes from my last two ventures. My motto for building an early startup team is lean and mean. And what I mean by lean and mean is hiring people that punch above their weight. I do not care about pedigree. I mostly care about people being honest

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

728.932

on board with the mission and willing to work hard and having a very tight knit of engineers that come to the office every day and have as high information sharing and collaborative environment as possible. And so if you ask me in a few words, what do I look for in an early hire? It is excitement.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

751.571

It is willingness to work hard and definitely low ego because it just makes the collaboration a whole lot easier. And so that will be like a short answer.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

871.498

I feel like when it comes to scaling machine learning models or architecture, We're still very much in the early days. I feel like most companies are still writing their own infra once they get to the stage where they need to scale either their training infrastructure or scale their inference infrastructure.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

893.125

There are tools built by big tech to make that simpler, but typically you run into limitations relatively quickly. And so, so far... I see most companies having to roll some version of the infra themselves just to fit their very specific use case. We plan to try to rely on kind of the cloud providers as long as possible. I don't want to build my own GPU cluster.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

923.016

I think that's a waste of time for most machine learning companies. So unless you're building an LLM that costs you 10 millions or more to train, I think you're very, you're much better off relying on someone else's infra and not take on that DevOps responsibility yourself. So we're definitely going to rely on the GPUs that we get from the cloud providers in the early days.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

948.441

And yeah, you know, you asked, have I seen the challenges? Yes, I've seen the challenges, but I feel like two years in the machine learning infraspace is an eternity. And so there will definitely be, you know, new frameworks that we'll have to use or will be new platforms that are probably going to be better pricing than they were a few years ago.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

968.669

Yeah, rely on your past experience, but approach every problem with a fresh mind and willing to be surprised by something new, better, making more sense than what made most sense two years ago.

Code Story

S10 Bonus: Heiki Riesenkampf, Commonbase

993.584

the optimizations that have gone to improve the latency of our very first product. Every customer call that we have, latency is always a question that comes up and it's always one of the main parts of the conversation. If you want to have real-time translation, it's always a question of, right, like, what is real-time? Is 2,000 milliseconds latency real-time? Is 1,000 milliseconds real-time?