"The data-readiness assessment surfaced three feature gaps in week one we'd been guessing at for a year. By week six we were running improved variants against our historical data."
Foundation ModelRead brief →
LLMs predict the next word. We train models that predict the next action — a behavioural world model for your institution, learned from transactions, trades, settlements, and event sequences. The same primitive serves fraud · recommendations · churn · LTV · trader intelligence · market microstructure · agentic decisions. A 12-week pilot, a forward-deployed delivery team, and artifacts you keep. Loom is the institutional deployment surface of ZKAI Labs.
"The data-readiness assessment surfaced three feature gaps in week one we'd been guessing at for a year. By week six we were running improved variants against our historical data."
"What we got is a model we own and a roadmap our team can extend — not a black-box vendor model. The joint model council made it impossible to drift away from the decisions that mattered."
"They bridge two worlds — behavioural intelligence and ZKP-aware attestation design — that almost no team can speak fluently. The architecture memo was usable verbatim with our investors."
The pilot delivers a production-grade behavioural world model trained on your event sequences — transactions, trades, settlements, on-chain actions, network signals — evaluated against your benchmarks, and shipped as artifacts you license. The same primitive serves fraud, churn, LTV, trader intelligence, market microstructure, or whatever use case your data calls for. We bring the methodology and the team. You keep the model.
A behavioural world model trained on your event sequences — the architectural pattern major payment labs publicly shipped on in 2025, applied here to your data across whatever use case it serves: fraud, recommendations, churn, LTV, trader intelligence, market microstructure, agentic decisions. One primitive, many heads. Delivered by a forward-deployed team you embed alongside yours.
Domain-specific transformer over your tokenized event sequences — transactions, trades, settlements, behavioural signals. A behavioural world model for your institution: the same primitive serves fraud, churn, LTV, trader intelligence, market microstructure, or recommendations depending on the head we train. Network-risk graph features use our belief-propagation methodology (validated in ZKAI Labs' moderation research). Optional on-chain features layered in where the data exists.
IP: You own the trained customer-specific model artifacts, code, task outputs, evaluation reports, and any new IP specific to your deliverable. Your data is yours in perpetuity.
The foundation model is the primitive. The heads are what your team actually ships: fine-tuned task layers on top of the same embedding space. We co-design and train the heads your team needs — and any one model can serve several simultaneously, with the costly pretraining shared once.
IP: Each task head is your IP — trained weights, code, task outputs. We retain general head-design patterns. New heads added later under a follow-on, not a re-build.
Three data-science FTEs · founder availability · research advisor · Embed advisor board. Weekly joint model council, fast-path data approvals, dedicated working sessions. Not staff augmentation — focused, scoped delivery.
Engagement: Includes implementation capacity to configure, adapt, evaluate, and deliver the agreed artifacts.
A pilot that ends with a production-ready model, evaluated artifacts, and a clear deployment roadmap — not a slide deck. Every week has named outputs reviewed by a joint model council with your team in the room.
Across your provided sources and selected on-chain / off-chain public sources. We catalog what exists, what's clean, what's labelled, and what's missing — the first honest map of your data surface.
Reproducible feature pipelines over your event data. Baseline behavioural model for the head we're targeting first (fraud, churn, recommendations, microstructure — whichever is highest-value). Network and risk-graph features using our belief-propagation methodology. On-chain features where the data exists.
Combine your data with off-chain and on-chain features. Backtest against your historical data. Model cards and evaluation reports against your domain benchmarks and universal benchmarks.
Production-grade model artifacts. Inference API or batch outputs. Recommendations for production deployment and (if applicable) protocol integration. Executive review of the production roadmap — the day-one inputs for an ongoing relationship.
Pilot success is evaluated against concrete criteria: successful ingestion of agreed data sources, production of feature schemas, completion of baseline and improved backtests, delivery of model cards and evaluation reports, and executive review of the production roadmap.
Trained weights, code, inference API, evaluation reports, and any new IP specific to your deliverable are yours in perpetuity. We retain general methodology and infrastructure patterns — never your model.
Training runs inside your environment or a dedicated tenant. Your data is never used to train another customer's model and is not retained after the pilot beyond what you opt into.
Twelve weeks ends with a model your team owns and operates. Production licensing and serving follow a separate agreement — opt in when you're ready, walk away when you're not.
The pattern: institutions with proprietary behavioural data — transactions, trades, settlements, customer interactions — wanting to compound it into a behavioural world model trained on their own data, not someone else's. Four shapes we've built for. The primitive is the same; the use case differs.
Institutions with proprietary customer behaviour data in the Global Majority — Mexico, Kenya, Philippines, India, Indonesia, Vietnam. The data exists. The next leg is turning it into a behavioural world model that can drive risk, fraud, churn, retention, and recommendations — and become an asset other institutions plug into.
Banks and trading desks where the existing ML stack is built for one segment but the actual business is everywhere else. We bring the event-sequence behavioural-model pattern — trade flow, market microstructure, execution quality, AML signal, trader intelligence — to the parts of the business that don't fit the legacy stack.
Fintechs whose first ML model fit one cohort and breaks at the next. The behavioural-model approach generalizes across cohorts because it pre-trains on behaviour, not labels — and Loom weaves fraud, churn, LTV, lookalike, and next-action heads on top of the same embedding space.
Aave, Compound, Polymarket, Hyperliquid-class protocols — the venue primitives are solved; the behavioural-intelligence layer above them isn't. Same primitive, different heads: counterparty risk, market microstructure, sybil resistance, recommendation, trader intelligence — depending on what the venue calls for.
We've shipped foundation-model-aware ML at consumer-app scale, published research that beats prior state of the art, and we operate natively on-chain — a combination that's rare in any team and rarer in a team you can hire for a 12-week pilot.
We built and operated a foundation-model-aware ML platform for blockchain personalization — battle-tested at scale across major crypto consumer apps. The same architectural pattern that major payment labs publicly shipped on for transaction modelling in 2025.
getembed.ai →Our network-classification research beats the prior state of the art on directed social-network sybil detection by modelling denouncement edges — blocks, reports, downvotes — which prior SOTA couldn't. The exact methodology we apply to risk-graph features in your behavioural world model.
zkai.network →Two years operating production ML across every major chain. ZKP-aware. Smart-contract literate. The bridge between off-chain behavioural data and on-chain markets — money markets, prediction, perps, DEX — is a load-bearing capability, and it's rare.
Backed by a16z crypto · Mask Network · Polymorphic Capital. Distributed team across continents. Founder availability is a feature of the engagement, not a marketing line.
A small team, founder-led, with the rare combination of foundation-model research literacy, production ML engineering at scale, and on-chain native fluency. Loom is our delivery methodology — the part that compresses what used to be a year of foundation-model work into twelve weeks.
Bring a clear use case, a real customer base, and a domain owner ready to sit on the joint model council. We'll bring the methodology, the team, and the pilot deliverables.