An AI-Native Trading Stack, Built From First Principles

When we left Deutsche Bank and Bank of America in 2022, we carried with us a clear thesis.

The electronic trading stack most banks and hedge funds rely on was built for a deterministic world: feed in explicit parameters, get a repeatable outcome. Quant work in this paradigm is largely a game of subtle model shifts and hyperparameter tweaks executed by a small army of highly paid quant specialists. Each team would carve out their sliver of microstructure, handcraft a signal, and then laboriously stitch everything together into a larger ensemble. When done well it prints, but it is still a slow and arduous process.

We weren't sure there was a better way, but we felt like advances in AI made it worth the risk to try something new. We also knew this risk was likely not something we would get permission to do at the bank. Not because people at the bank are not smart, but because they are careful risk managers. That is precisely how you stay alive. When you are sitting on a business that prints cash every year you don't blow it up for an unproven technology like AI, and keep in mind this was 2022.

On the other hand, if you wanted to take the risk, then we felt like AI was a once-in-a-generation opportunity to reimagine the entire trading process from the ground up. Instead of adding more people, we could build systems that learn market dynamics holistically, compress months of traditional quant labor into overnight runs, and treat AI models as first-class citizens rather than afterthoughts bolted onto legacy infrastructure.

We named the company Deep Q as an homage to the Deep Q reinforcement learning algorithm. Digital asset markets became our laboratory. Not only did we believe that digital assets offered a lower barrier to entry than traditional asset classes, but we also believed that they would play a pivotal role in reshaping the efficiency and transparency of capital markets.

The best quote I can think of to describe our journey thus far is, “We do these things not because they are easy, but because we thought they were going to be easy.”

We have made a lot of false assumptions about both AI and digital asset markets over the last few years, but we also solved some really gnarly problems and have built, and put into production, a complete AI-native trading stack. The four core systems are Prometheus, Metis, Hive, and Pebble.

Prometheus

Generating consistent alpha in electronic trading is difficult. Markets are efficient. If you quote naively without a view, informed participants will run you over. The traditional solution was teams of quants hunting for subtle patterns in order-book dynamics and combining their findings into an ensemble model. That approach works, but it is expensive, slow to compound, and typically scales with headcount. When you are a startup you have to find alternative solutions.

Prometheus takes a different path. We built a neural-net architecture that learns a compressed, multi-market representation of order-book dynamics. Drawing on techniques from modern LLMs and world models, but applied to the lower-dimensional and considerably noisier world of tick-level market data, it predicts sequences of future order-book states across assets and venues simultaneously. The institutional expertise we developed over years on trading floors is still essential here. It guides how we structure the model, what noise to filter, and where purely data-driven approaches would go astray.

We were skeptical about neural nets as a source of alpha for all of the reasons quants should be skeptical about neural nets, but through lots of trial and error we found a way to make it work in a scalable and stable way. Not only that, but because it predicts sequences of full orderbook states rather than just a singular forward price prediction, it can be extended to a host of other interesting problems besides simple price prediction.

Metis

Even with strong predictive models, research remains the bottleneck. Backtesting, data validation, post-trade analysis, and strategy iteration are still largely manual work that requires a lot of context-switching. A researcher spends an enormous share of their time managing the mechanics of the process rather than thinking clearly about the market structure problem in front of them.

Metis was built to handle that mechanics layer. Using the latest generation of large language models inside a custom coding harness, it acts as an always-on research collaborator with deep knowledge of our trading systems, historical data, backtesting standards, and data-quality protocols. All of that institutional knowledge is encoded as skills and tool calls that Metis can draw on autonomously. It can run complete research cycles overnight while the team sleeps.

One thing worth saying plainly: this is not what you get by prompting Claude Code to “become a quant research assistant, make no mistakes”. Metis knows our environment intimately, follows our research standards, and operates under strict guardrails that prevent it from moving anything into production without human review. There is no shortcut to that layer. It required putting in the reps to understand both the technology and the domain it operates in.

Hive

Prometheus provides directionality. Metis accelerates research. Hive ties everything together.

It is the orchestration layer, and the best way to think about it is as a portfolio manager that operates through reinforcement learning. Any trading strategy can be reframed as a classic reinforcement learning game: what is the state of the market I am observing, what actions can I take, and what reward am I likely to receive conditional on taking a given action. Hive converts strategies into that structure and then searches across the space of possible decisions in a way that traditional grid search and manual hyperparameter tuning struggle to replicate.

Reinforcement learning is notoriously unstable, it has a hard time extrapolating outside of the bounds of data it has already seen, and it bears repeating that financial data is extremely noisy. For this reason much of reinforcement learning's early successes came from solving stable problems where you could clearly map a reward to a specific action. But what about trading where the same action taken in seemingly identical circumstances can yield wildly different results. In spite of this, somehow the best traders in the world still learn behaviors that allow them to be consistently profitable over the long run even if they are wrong on a high percentage of individual trades. The key learning was that teaching the model winning behaviors in the face of uncertainty was more powerful than the model trying to be correct on an individual trade.

But why solve these problems in the first place? Firstly, because we thought it would be easy. Secondly, because when you solve these problems you get a system that not only trades profitably, but also has the ability to adapt and improve upon itself, eventually finding novel solutions to problems that would normally take multiple quants multiple months to solve.

Pebble

None of the above works without a foundation purpose-built for it.

Legacy trading systems were designed for human traders and deterministic models. They were built to be fast at a specific kind of computation, not flexible enough to serve as the substrate for probabilistic AI outputs running across a dynamic mix of on-chain and centralized venues. Retrofitting AI onto that kind of stack is the equivalent of bolting a robot to the steering wheel of a car and calling it an autonomous vehicle. Technically it works, but you really have a robot chauffeur.

We built Pebble from scratch. Think of it as reimagining the car's interior from scratch, knowing the model would be doing all of the driving. Certain components that traditional trading infrastructure assumed were necessary simply become irrelevant. Others that never needed to exist before become critical. Pebble runs on low-cost commodity cloud hardware, handles both on-chain and off-chain execution, and was designed from day one so that AI models are native throughout the stack rather than connected to it.

Why This Matters for Traditional Markets

We did not build this stack solely to chase retail crypto volume. We built it because we saw the same structural problems that still exist across traditional capital markets. Quant compensation continues to climb. Annual infrastructure budgets run into the tens of millions. And the growing realization spreading across the industry is that handing teams access to the latest AI tools will not automatically resolve decades of accumulated technical debt. In reality, you need purpose-built models for financial data. You also need the expertise and restraint to build systems that prevent the models from making catastrophic mistakes.

The combination of deep institutional experience and ground-up AI-native engineering is rare. Teams that have only ever worked in traditional finance are just beginning to experiment with these tools. Teams that are purely AI-native often lack the hard-won intuition about how adversarial markets actually work. We have lived in both worlds, and we built this stack at the intersection of them.

Today our systems are live, trading spot, perpetual futures, and prediction markets. As blockchain infrastructure continues to blur the lines between digital and traditional assets, this framework extends naturally across asset classes.

If this message resonates with you and you would like to learn more about what we do, we would welcome the conversation.