An applied-AI research lab for finance and on-chain markets. Embed is our production credential. Moderation at scale is our research credential — beats the prior state of the art on directed social-network sybil detection by modelling adversarial edges the prior couldn't. Next: foundation world models for markets — pre-trained over event sequences, with learned simulators built in. Built for the moment finance defaults on-chain — agent-economic, stablecoin-settled, network-effects-led.
Five applied capabilities. Production-grade. Built on a shared infrastructure that runs across every major chain and venue. Each one validated at scale before it leaves the lab.
Foundation-model-grade ranking. Behaviour-aware feeds, alerts, and agent-callable endpoints — the surfaces that matter when most decisions are made by software, not humans. The substrate we already built and validated under real-world adversarial load.
Sequence models trained on adversarial behavioural data — on-chain and off. Detection at the boundary where bad actors evolve fastest and labels are sparse.
Engagement-intelligence models for the high-churn surfaces where retention is the constraint and the signal-to-noise ratio is low.
Event-sequence foundation models over orderbook, trade, and on-chain data. Multi-horizon. Multi-asset. Calibrated for production trading desks.
Agent-based markets calibrated against real behavioural data. Stress test fees, listings, incentives, and microstructure designs before they touch production. The same machinery underpins the world models we're building next.
The ML platform we built for blockchain personalization. Foundation-model-aware. Shipped at scale. The credential behind what comes next.
A foundation-model-aware platform. Ingests on-chain events as tokenized sequences, learns wallet-level behavioural embeddings, serves personalization across feeds, alerts, leaderboards, and agent endpoints — the same architectural pattern major payment labs adopted for off-chain transaction modelling in 2025, applied to on-chain behavior first.
The platform is the proof of what comes next.
Applied research is the other half of what we do. Moderation at scale is the first output out of the lab — a network classifier that beats the prior state of the art by reading signals the prior couldn't see. Spam classification is one of half a dozen applications. The framework is generic.
The problem. Networks need to know who's a spammer, who's a sybil, what's a bot ring — from signals that are noisy, sparse, and adversarial. Local features alone aren't enough. Network structure carries most of the information.
The approach. The model treats the network as a probabilistic graphical model — a Markov Random Field over the social graph. Local features give noisy priors per node; belief propagation across both endorsements (follows, upvotes) and denouncements (blocks, reports) upgrades those priors into much sharper posteriors. Bayesian belief flow at network scale.
The result. On the Farcaster network — with only user-user relations, deliberately low-quality priors, and zero hyperparameter tuning — the MVP demonstrated strong results out of the box. The prior state of the art — SybilHP (Lu et al., 2023) — adapts to directed graphs with iteratively-estimated edge homophily, but treats every edge as an endorsement. It can't see the most informative signal in an adversarial network: who's getting blocked, reported, or downvoted. Ours can. That single asymmetry is the win.
The framework. V0 ships a generic entity-opinion graph applicable across protocols. V1 (under design) extends to special entity classes and bipartite graphs, configurable per integration — so new protocols slot in without redesign.
Internal repository · paper in preparation · partner access on request
The substrate is the same: tokenized event sequences, behavioural embeddings, multi-horizon objectives. The surface changes. Three areas where the work lands — and where the next billion-user crypto wedge actually shows up.
Underwriting, fraud, default prediction, cross-sell — recalibrated for a world where credit is denominated in stablecoins and rails settle in seconds. The clean ML problems where a 2-point lift on AUC is a quarter of revenue.
Personalized discovery, smart-money intelligence, market-making signals, listing impact simulation — and the agent-callable layer that becomes load-bearing when most economic action is initiated by software, not humans.
Multi-horizon forecasting, tokenized-RWA discovery, on-chain signal extraction for off-chain books. The model that bridges TradFi to on-chain markets in the cycle when equities, treasuries, and credit start trading on-chain by default.
Two research directions, both load-bearing on the next decade of AI infrastructure. Foundation models for finance is the natural extension of Embed. Compute markets are the financial primitive that makes it affordable.
Transformers over tokenized event sequences — blockchain transactions, orderbook microstructure, and behavioural signals. Pre-trained on the corpus that came out of Embed. The architectural pattern the major payment labs publicly shipped on in 2025 for off-chain transactions; the on-chain analog hasn't been built yet. Privacy-aware where it counts — encrypted user contexts the model can serve without leaking, because in a world where block space is commoditizing, that's the durable moat.
The thesis: text-trained LLMs aren't built for finance. The right primitive is a model native to event sequences — what wallets did, in what order, across which venues, at what timestamps. JEPA-style pretraining: predict masked event embeddings from context, not reconstruct raw tokens. The same family of architectures LeCun's lab used this year to control real robot arms from passive video, applied here to wallet trajectories — and to the AI agents that will initiate most on-chain transactions before this decade is out.
If foundation models are the next decade's defining workload, GPU access is the defining financial primitive — the thinking sand that produces them, plausibly the most important asset of the next era. Decentralized GPU networks — Akash, io.net, Render, Aethir, Fluence — already aggregate hundreds of thousands of consumer and enterprise chips. The market layer above them does not yet exist: hedge instruments for compute capacity, settlement rails between centralized and decentralized supply, price discovery for inference at scale.
The bet: compute will trade like electricity does today — futures, options, capacity swaps. Decentralized providers become liquidity. Frontier labs become hedgers. And capital markets for compute will be built on-chain by default — because new markets in this cycle are. The GPU compute market, projected $83B in 2025 to $350B+ by 2030, becomes the financial primitive we build microstructure for.
Distributed across continents. Foundation-model research depth, production ML engineering, and on-chain native fluency — backgrounds that don't usually sit in the same room, working asynchronously.
Research collaborations, capital conversations, strategic partnerships. ZKAI Labs is a small applied-AI research team in Singapore — the door is open to peers, investors, and partners working in the same space.