Building Artificial Intelligence

I am a non-executive director and ex COO for Tyler Capital Ltd (TCL), a FinTech-based hedge fund based in London. I consult for and help other companies which aim to leverage technology. At TCL we focus 100% on machine learning and neural networks (so-…

I am a non-executive director and ex COO for Tyler Capital Ltd (TCL), a FinTech-based hedge fund based in London. I consult for and help other companies which aim to leverage technology. At TCL we focus 100% on machine learning and neural networks (so-called ‘deep learning’) to trade financial markets.

Writing in a personal capacity, I share how we successfully implemented machine learning. Please do get in touch if this is of interest and if you’d like to share your thoughts, business ideas or otherwise.

So how did we build a successful AI business?

1. Core purpose — an exclusive focus on deep learning, an ambitious goal

We decided 5 years ago that neural networks and machine learning (ML) would become our core purpose to the exclusion of other methods and approaches. Previously we had tried to ‘bolt on’ a new siloed ML project to existing systems and methods and while this perhaps was necessary as a proof of concept for developing the neural net, we quickly realised we had to focus and invest. We made some tough decisions and restructured the business around the core focus of ML. Our size helped (we’ve never been more than 75 people), flexibility and agility have been crucial.

Over the 5 years, we invested over $50m dollars in building a unified platform, admittedly making multiple mistakes along the way but doing AI properly is not simple and does not come cheap. We had an ambitious goal, to ‘create the world’s greatest trader’.

Today we have achieved the ability to generate consistent profits exclusively through an ML platform. We have built something truly innovative.

2. Strategy, plan and business model: a production line for innovation

We had a clear plan, but it evolved and continues to evolve.

Systems thinking is core to what we do. We think of the organisation itself as a unique, complex, unified, single system.

We focus with laser-like precision on process and we built a business model around the manufacturing of ML models, with concepts taken from the industrial production line. Basically, this means continual, rapid and structured prototyping, simulation and deployment of AI models. Controlled failure is valued with deep analysis, but lessons are learned quickly to allow us to advance quickly.

Process is what drives innovation.

3. People and culture; humans with a core role in the ML

We overhauled our team and recruited some of the brightest and the best, from the very top downwards. Nevertheless, our model means we don’t need a genius creating ‘magic’, our team has no single star and can survive staff turnover, it is resilient.

Like many successful companies, we recognised that success needed good people working in a strong and supportive culture. This is highly respectful but also entrepreneurial. We instilled a ‘servant-leader’ management approach, where our seniors look to serve as much as to lead. Our staff is, as a consequent, deeply empowered. Talented people want to come and work for us when they see this is the reality, although they must be a good ‘fit’ meaning we will turn down characters that don’t.

Indeed, important in our core ML approach is that humans have a key role engaging directly with the platform. We are pure ML, but even that requires human interaction. ML lacks context and that at times can be a serious drawback. Balancing this, without doubt, remains a major challenge in ML.

4. Openness and ethics

We removed all internal silos and prioritised openness and communication. We accept the security risks this entails (any developer can access our core code base for example) because openness fosters progress. (N.B. we do take security very seriously — our security model is likened to an armadillo, very hard exterior but soft interior).

We prioritised AI ethics, governance and ‘explainability’ (AIX). We built a whole suite of systems, tools and processes that enabled us to best understand our AI and ensure it prioritised, (as much as we can be sure — a key challenge with AI), the best regulatory and governance standards. We’ve been patient and persistent and have a very healthy view of managing our risk.

5. Reliability and resilience

In trading financial markets, reliability is a real concern, matters can go wrong quickly so as a result we became a ‘High Reliability Organisation’ (HRO). This is a concept applicable previously to power stations or aircraft carriers where a systems failure might be catastrophic. It embeds utmost respect for quality and promotes resilience, ultimately aiming for zero defects. Testing is fully integrated and automated, (we employ no ‘testers’ as such). Our desire to analyse and understand our processes and performance, notably when it is imperfect or just goes wrong, has become an obsession. We have been the focus of a study on this published by the University of Copenhagen.

6. Data

Excellent and bountiful data is the lifeblood of ML. Without it, there can be no successful ML implementation. Working in financial markets, data, while expensive, is abundant and yet significant time and effort is still spent ‘cleaning’ and preparing it (during which we must be careful not to add bias). This data, the processes and infrastructure around it, is, after salaries, our single biggest expense. Getting data right is without doubt a massive and crucial undertaking. We went through multiple data vendors, systems and data specialists and ultimately decided to build our own solutions from scratch (although we still have to buy in a lot of data itself). This is just not something that can easily be outsourced. Data is ultimately everything.

7. Raw power

There are many outsourcing options for firms looking to ML solutions, these can all help speed development. Ultimately however, we ended up building nearly everything ourselves from the core neural network to data capture and storage systems and all analysis, risk management and compliance monitoring. What can be facilitated externally however is raw computing power. Cloud service providers have become extremely efficient at providing this power when you need it. The widespread availability of cloud computing services empowered our model. It is expensive but now ubiquitous and so a low-risk choice to outsource.

Mastering ML is not some project that can be developed ad hoc with limited resources by a small team that’s locked away apart from the core organisation. It is daunting but there are methods and techniques that are applicable and can be taught. You start with leadership focus, skills and experience.

I will be the first to admit, as with any business, there have been many bumps along the way, running ops means implementing strategy and driving the business forward but also dealing with a multitude of challenges, from legal and regulatory, technology and systems to people and culture that just must be dealt with. Our business remains a work in progress.


Building Artificial Intelligence was originally published in Level Up Coding on Medium, where people are continuing the conversation by highlighting and responding to this story.


Print Share Comment Cite Upload Translate
APA
Paul Tyler | Sciencx (2024-03-28T13:54:24+00:00) » Building Artificial Intelligence. Retrieved from https://www.scien.cx/2021/02/26/building-artificial-intelligence/.
MLA
" » Building Artificial Intelligence." Paul Tyler | Sciencx - Friday February 26, 2021, https://www.scien.cx/2021/02/26/building-artificial-intelligence/
HARVARD
Paul Tyler | Sciencx Friday February 26, 2021 » Building Artificial Intelligence., viewed 2024-03-28T13:54:24+00:00,<https://www.scien.cx/2021/02/26/building-artificial-intelligence/>
VANCOUVER
Paul Tyler | Sciencx - » Building Artificial Intelligence. [Internet]. [Accessed 2024-03-28T13:54:24+00:00]. Available from: https://www.scien.cx/2021/02/26/building-artificial-intelligence/
CHICAGO
" » Building Artificial Intelligence." Paul Tyler | Sciencx - Accessed 2024-03-28T13:54:24+00:00. https://www.scien.cx/2021/02/26/building-artificial-intelligence/
IEEE
" » Building Artificial Intelligence." Paul Tyler | Sciencx [Online]. Available: https://www.scien.cx/2021/02/26/building-artificial-intelligence/. [Accessed: 2024-03-28T13:54:24+00:00]
rf:citation
» Building Artificial Intelligence | Paul Tyler | Sciencx | https://www.scien.cx/2021/02/26/building-artificial-intelligence/ | 2024-03-28T13:54:24+00:00
https://github.com/addpipe/simple-recorderjs-demo