At Dexterity Capital, testing a trading idea once meant a person sitting with the data, forming a hunch about which signal predicted which price move and checking it, maybe once a week or once a month. Abraham Chaibi, who co-founded the crypto trading firm in 2017, turned that loop into something a machine ran thousands of times a day.
He calls the result a brain. It’s a slightly playful word for what was, physically, a room full of servers, but it fits what the system did. It looked at enormous amounts of market data, figured out which pieces predicted which events and turned those findings into strategies. “Effectively, what the brain was, was it was a giant data center that would process all of this and come up with algorithms that would make money,” Chaibi said.
Abraham Chaibi didn’t start with a data center, though. Dexterity’s first edge came from a single smart contract and a price gap.
$30,000 and a spread
Chaibi and his co-founder, Michael Safai, started Dexterity with about $30,000. By 2018 they had turned it into roughly $2 million, according to a 2022 profile of the firm in Decential Media.
The opening came from how disorganized crypto markets were in 2017. Early DEXes like Bancor and EtherDelta let anyone trade ERC-20 tokens peer to peer, and the same token often carried different prices on different venues at the same moment. Chaibi wrote a smart contract that could run several steps in one transaction and refuse any trade that would have lost money. If buying on one exchange and selling on another cleared a profit, it went through. If not, nothing happened.
“It was really fun back then, we did lots of cool things that I just couldn’t believe worked,” Safai told Decential.
He also found a quieter edge inside Ethereum itself. Every transaction on the network costs “gas,” and the price of gas swings with demand. At the time, the network paid back part of the fee when a contract cleared out stored data. Chaibi optimized Dexterity’s contracts and built an early gas-token arbitrage around SSTORE, the instruction that writes to contract storage. The general idea behind gas tokens was to buy cheap gas when the network was quiet and use it when fees spiked.
Neither trick would last forever. Spreads between exchanges close as more traders chase them, and Ethereum’s developers later changed the refund rules. What lasted was the habit of writing code that exploits an inefficiency safely and then letting it go once the gap closes.
From clever contracts to a research machine
Chaibi came to trading with a background in optimization and automation. He studied mechanical engineering at Princeton, worked on NFL scheduling and retail inventory at the Boston firm Analytics Operations Engineering, and automated the movement of petabytes of files as a software engineer at LiveRamp in San Francisco. Every one of those jobs involved looking at something people did by hand and asking whether code could do it better.
A trading strategy is a hypothesis about the future, and hypotheses can be tested. At Dexterity, the limit on progress was how fast those tests could run.
That was the problem the brain solved. Chaibi described the goal as building “a procedure by which we could figure out which bits of data predicted which events,” with the whole thing running automatically. What had been a person reasoning through an idea once a week or once a month became a pipeline that could run “thousands of times a day.”
Running that many experiments takes a lot of computing power, and renting it gets expensive fast. Dexterity built its own data center instead, with 10,000 processor cores, and cut the total cost of fitting models by 90% compared with cloud providers. For a firm whose edge depended on testing more ideas than its competitors, that saving went straight into more research.
What 100,000 trades a day looks like
By 2022, Decential reported, Dexterity was making between 100,000 and 200,000 trades a day using market-neutral, high-frequency strategies. Market-neutral means the firm wasn’t betting on whether crypto prices would rise or fall. It was trying to profit from small, statistically reliable patterns in how prices moved relative to each other.
There was a great team behind the algorithms. He wrote upward of 100,000 lines of it, tuned for speed, and over his time as CTO it moved more than $2 trillion in volume. The firm ran $300 million across dozens of exchanges with reaction times measured in milliseconds. To staff it, Dexterity recruited more than 30 engineers and traders, some of them from Hudson River Trading, Radix and Tower Research.
At that speed, the exchanges themselves become part of the problem. In March 2023, as FTX’s bankruptcy team floated the idea of restarting the collapsed exchange, CoinDesk went looking for speed-sensitive former customers who could say how well its technology had actually worked. Dexterity was one of them, and Chaibi gave one of the more concrete technical assessments in the piece. Round-trip latency on FTX was typically about 150 milliseconds and rose to 600 to 800 milliseconds when markets got busy, he said, compared with 5 to 10 milliseconds on Binance. Fill notifications were so slow that “if you actually wanted to know promptly that your order had been filled you needed to repetitively query the state of your order,” checking every millisecond.
He was blunter about FTX’s collateral rules. The exchange let customers withdraw dollars against balances of other coins, including its own FTT token, with very little discount. “No other exchange lets you withdraw a negative balance like this,” he told CoinDesk. “It is and was insane.”
Reporters also came to him when markets did something strange. On September 15, 2022, Ethereum completed the Merge, its long-planned switch to proof of stake, without any technical hitches. Ether fell anyway, down 9.1% to $1,489 by CoinDesk’s press time, its biggest one-day drop since late August. Chaibi’s explanation was about positioning. In his view, plenty of funds were holding large hedged positions tied to the event, and those trades would come off once it was over. “I think that you wouldn’t have to throw many rocks to find a fund that has a healthy nine-figure position on the Merge, long spot short derivatives or something like that,” he said. “They’re going to unwind, that’s definitely going to happen.”
The last 20% is where the lessons live
Ask Chaibi what was hardest to automate in trading and he doesn’t talk about models or latency. He talks about everything that gets added after a system goes live.
“You kind of build, like, the first 80%, and then there’s a bunch of additional code that comes in over time as you learn from experience,” he said. He described how it builds up. “This thing happened, and then we lost a lot of money, so we added in this protection. This thing happened, and then it was a disaster, so we added in this protection.” Over time the code fills up with safeguards, each tied to a specific thing that went wrong.
His comparison is a bicycle. On the first day you can get on and roll down a hill. Everything after that comes one experience at a time, and he describes those lessons as “very specific and minutiae.” You can’t design them in advance, and a mature system is largely made of them.
For Abraham Chaibi, the brain and the safeguards were two halves of one idea. A system that runs thousands of experiments a day runs into its weak spots sooner, and each one becomes another protection in the code.
Carrying the brain into a lab
After eight years at Dexterity, Chaibi moved on in late 2024. Today he is co-founder and CEO of Sapho Bio in San Carlos, California, which runs rapid sterility and other release tests for pharmaceutical products. On paper it’s a big change. Market microstructure has little to do with bacteria.
He sees more overlap than that. Sapho meets similar testing problems over and over, he said, and every solved one goes into a shared body of knowledge that makes the next one quicker. It’s the same pattern as the trading code, where every hard lesson became a permanent part of the system.
The bigger difference is that the code now drives physical equipment. Early in his career, automation meant pure software, since equipment in the physical world was so difficult to control. That is no longer true, he said, and now “the software can do things.”
He still builds things for fun, too. One example he gave is a personal wiki, powered by language models, that keeps track of everything he is working on and answers questions about it. It’s a small, one-person version of the Dexterity brain, a system that holds what he has learned and hands it back when he needs it.
The first version of all this, in 2017, was a single contract built to refuse a losing trade. His current company checks drug batches for contamination before they ship, a job where the system also has one outcome it must never allow.
