attimet wants its models to learn the market, not memorize it

The Y Combinator directory is full of elaborate one-line pitches. attimet's reads, in full: "we predict stuff."

That is either the least informative tagline in the batch or the most precise, depending on how seriously you take it. Take it seriously. Prediction, stripped of everything else, is the entire business of quantitative trading. Every signal a quant firm hand-crafts and every microwave tower it builds between Chicago and New Jersey is in service of one question: what happens next, slightly before everyone else knows it?

attimet is a San Francisco research lab betting that the best answer to that question is no longer a hand-built signal. It builds temporal AI systems, models designed to ingest live market data, adapt as conditions change, and act in real time, and it tests them the only way that counts: in live trading experiments, starting with options. The company came out of Y Combinator's Fall 2024 batch and closed a $10 million seed round in December 2024, per PitchBook, with backing from Neer Venture Partners, Orange Collective, and YC itself, bringing its total raised to about $10.5 million. That is an unusual amount of conviction for a company that was a few months old, until you look at who started it and what they left behind to do it.

Two Optiver alumni pick the hard market first

attimet was founded in 2024 by Kirthi Banothu and Xiaoyu Li, and their backgrounds explain most of the company's design choices.

Both came through Optiver, one of the world's premier options market-making firms. Li was a quant trader there, specializing in index and equity options market-making, and later started and grew a medium-frequency algorithmic trading team. Banothu built low-latency trading systems at Optiver and DRW, developing strategies across futures, crypto, equities, and options. In between, he led the onboard sensor and ML libraries team at Argo AI, the self-driving company, which turns out to be surprisingly relevant experience: a self-driving car is also a machine that must perceive a fast-changing environment, predict what happens next, and act on that prediction in milliseconds, with real consequences for being wrong.

So when this particular pair chooses to start with options, it is worth noticing. Options are the market they know best, and they describe the space the way practitioners do: complex, under-modeled, and full of opportunity. An options market is not one price to predict but a whole surface of them, thousands of strikes and expirations per underlying, each repricing continuously against volatility, interest rates, and the movement of everything else. The combinatorial sprawl that makes options brutal for hand-built models is exactly what makes them attractive terrain for models that learn structure from data.

The founders are not tourists discovering that markets are hard. They spent a decade inside the firms that dominate this game, and they left to attack it differently.

Learning the market instead of hand-crafting it

The traditional quant workflow has a shape that has held for decades. Researchers form a hypothesis about market behavior, encode it as a signal, backtest it, argue about it, and eventually wire it into a trading system. The intuition comes from humans; the machine executes it. Even at firms drowning in data, the signals themselves are largely artisanal.

attimet's approach inverts that. The company describes itself as taking a first-principles approach to trading: designing systems that learn and adapt with data rather than relying on hand-crafted signals and human intuition. Its research spans predictive models, reinforcement learning agents, and signal discovery workflows, working with structured and unstructured data at scale. The word the company keeps using is "temporal," and it is doing real work. A market is a stream, not a static dataset to be fit once, and a model that cannot update as the stream shifts is a model that decays. attimet is building systems meant to adapt continuously, closer in spirit to the perception stacks of autonomous vehicles than to a backtested spreadsheet formula.

There is a reason this is being attempted now rather than ten years ago. The machine learning techniques that transformed language and vision, sequence models, representation learning, agents trained against feedback, have matured enough to be pointed at domains that punish sloppiness. Markets punish sloppiness faster than almost anything. That is precisely what makes them an honest proving ground: the feedback arrives in minutes and is impossible to argue with.

The edge is iteration speed, not microwaves

The most revealing line in attimet's own material is about what its edge is not. It is not secret data, and it is not microwaves. The edge is the speed at which the team iterates and learns.

That sentence carries a quiet critique of where high-frequency trading ended up. For years, the marginal dollar in fast trading went into physics: microwave towers, shortwave links across oceans, cables laid a few miles straighter than the competition's, all to shave microseconds off the time it takes to see a price and react. That race was real, and the founders ran in it. It is also largely spent. When everyone's light travels in nearly straight lines, being fast is table stakes rather than an advantage.

attimet is relocating the race. If the latency of signals has been arbitraged to the floor, the next competition is the latency of learning: how quickly a firm can form a hypothesis, train a model, test it against reality, and fold what it learned into the next attempt. A firm that completes that loop in days while incumbents take quarters compounds its advantage the same way fast-shipping software companies compound theirs.

This is why attimet builds infrastructure with the seriousness most startups reserve for product. The company's systems ingest market data, train models, simulate strategies, and run them live, with feature stores, experiment tracking, and monitoring designed so that researchers move fast and every experiment feeds the next one. The point of all that plumbing is a feedback loop that connects research directly to market outcomes. An idea ends its life in a P&L rather than in a paper or a slide, and the result flows back into the research queue.

A lab where the market grades the homework

Calling yourself a research lab has become fashionable, and it usually signals long horizons and distant payoffs. attimet's version is stranger and more disciplined: a lab whose experiments settle in the market, sometimes the same week they were conceived.

Researchers there integrate alternative data, design experiments, and test hypotheses in simulation and then in live trading, working directly with the founders. The company stays deliberately small while doing it, a handful of people rather than a trading floor. That is only possible because the model of the firm is different. A traditional quant shop scales by adding traders and desks. A firm whose strategies are learned rather than hand-built scales by improving the machine that produces strategies, and a small team with the right infrastructure can run an enormous number of experiments.

One hiring detail says a lot about the culture. attimet does not require financial markets experience for its research roles, and says so explicitly: drive and curiosity outrank domain pedigree. From founders this steeped in markets, that is a considered position. The scarce skill for what they are building is knowing how to make models learn from streams of messy data, whether those streams come from lidar or the options tape. Markets knowledge is the part the founders can teach; a decade of Optiver and DRW between them is the domain-expertise backstop that makes hiring pure researchers safe.

The quiet reinvention of the prop firm

Step back and attimet fits a pattern worth watching: the trading firm rebuilt as an AI-native startup.

The great quant firms of the last generation, Renaissance, Jane Street, Citadel Securities, Optiver, are extraordinary businesses, and they were all architected before modern deep learning existed. Machine learning arrived at those firms as a tool bolted onto an existing process. attimet belongs to a newer species, built venture-backed and from scratch on the premise that learning systems are the process, with everything else, infrastructure, team shape, even hiring, derived from that premise.

Venture capital has historically kept its distance from trading strategies, preferring businesses that sell software to businesses that take positions. A $10 million seed for a months-old trading lab suggests that calculus is shifting. If a small team can build systems that genuinely learn markets, the prize is a compounding machine in one of the largest and most liquid arenas on earth, not a SaaS revenue line, and options are just the opening market. The same temporal models that learn an options surface have obvious neighbors: futures, equities, crypto, anywhere the data is rich and the feedback is fast.

None of this makes the problem easy. Markets are adversarial in a way vision and language are not; they adapt to the people predicting them, and every fund that ever blew up had a backtest that looked wonderful. The honest version of the pitch is that attimet is running a live experiment about where trading's next edge comes from, with its own capital markets experience as the prior.

But the experiment is well designed. Two founders who mastered the old edge are betting the next one belongs to whoever learns fastest, and they have built the company as a machine for learning fast. "We predict stuff" reads like a joke until you realize it is a scope statement. Prediction is the whole game. attimet is just refusing to pretend the game is about anything else.