ZKAI Labs is a small research and engineering team building behavioural world models for financial and agentic systems. $3M pre-seed, backed by a16z crypto CSX, Mask Network, and Polymorphic Capital. The team stays small on purpose — the people we'll want to work with are usually people we've already been talking to for a while. This page is where that conversation starts.
LLMs predict the next word. We train models that predict the next action — behavioural world models for institutions and agentic systems, learned from event sequences, on-chain transactions, social graphs, and agent interactions. Loom is the institutional deployment surface. Embed is the productized crypto model. The lab sits behind both.
No live listings — we open seats deliberately, one shape at a time. These are the shapes that show up most often when we do. If you read one of them and think "that's me," that's the message we want.
People who can read a transformer paper on Friday and reproduce the result on Monday. Comfortable with self-supervised pretraining, tokenization of non-text sequences, graph features, and the literature around behavioural foundation models (payment-lab papers, Markov random fields, belief propagation). Production literacy matters — this is not pure-research.
Active interestThe pilot delivery shape: someone who can sit alongside a customer team, ingest messy data, ship feature pipelines, train baselines, and stand up an inference API — end to end — in twelve weeks. Comfortable in Python and TypeScript. Reads code faster than they write Slack messages. Cares about evaluation as much as architecture.
Always-openQuantitative depth on real institutional data — risk, fraud, churn, market microstructure, on-chain behaviour. Strong intuition for what a feature is doing and what a model is missing. Can run an honest backtest, write a model card, and explain a calibration curve to a non-technical risk owner.
Active interestSmart-contract literate, ZKP-aware, and able to translate behavioural model outputs into on-chain attestations that prediction venues, lending markets, and perp DEXs can consume. Rare combination — the bridge between off-chain ML and on-chain markets is a load-bearing piece of the thesis.
OpportunisticThe team stays small for a while. That means real ownership from week one, direct lines to the founders, and the chance to shape the work as much as deliver it. These are the things we care about most — and what we look for in the people we end up working with for years.
You read papers and you ship code. The combination is what lets a team this size deliver foundation-model work in twelve weeks — and what makes the work fun. Researchers who like building, engineers who like the literature: this place is built for both at once.
Most of what gets published in ML doesn't reproduce. Most of what gets demoed doesn't work in production. We love working with people who'll say "the AUC is 0.71 and that's worse than the baseline" — because that's the conversation that gets us to v2. Honesty here is a feature, not a friction.
The roadmap shifts. The thesis sharpens. The customer's data turns out to be dirtier than anyone admitted. People who love it here treat all of that as the interesting part — because that's where the actual learning happens, and most of the leverage too.
Singapore HQ, team across continents. Async-first writing, a weekly sync rhythm, and the kind of trust that doesn't depend on a Monday standup to know what matters this week. Work from wherever you do your best thinking — we'll meet you there.
We read every email. A good intro is two short paragraphs, a link to your work, and one specific thing about the thesis that resonated — or that you disagree with. We're more interested in a thoughtful disagreement than a polished resume.
The team is small enough that the founders read every introduction. We try to reply within a week, even when there's no open seat — because by the time there is one, we want to already know who we're calling.