Mobile application behavior datasets
Source-level app behaviour, permissions, endpoints, consent timing, and observed data flows.
Data Solutions
production, data and agents that have to work
The same discipline behind our systems and public evidence, delivered as data and infrastructure: a data warehouse, reproducible environments, expert trajectories, and evaluations for difficult, long-horizon work. Our Rust and Axum services connect through a SQL bridge to ternary-logic pipelines. Your corpus stays private. Every delivered artefact stays traceable.
SDK & tracker knowledge graphs
Reusable SDKs, tracker relationships, infrastructure, and cross-application recurrence mapped as connected evidence.
Agent safety trajectories
Tool-use traces under injection, conflicting instructions, boundary pressure, recovery, and stop conditions.
World-state & causality mapping
Multi-domain observations, entity relationships, change histories, and evidence-linked causal chains.
Stressed benchmark variants
Controlled, reproducible perturbations of open benchmarks for robustness testing beyond clean inputs.
What we deliver
From environment to evaluation
We define success criteria with your team, then build the data operation around the way your models are actually trained and evaluated. No anonymous task stream and no benchmark theatre: the environment, trace, label, and decision remain connected.
Data warehouse
A governed foundation for operational, training, and evaluation data, connected to our Rust/Axum services through a production SQL bridge.
Virtual environments
Human-simulated companies, computer-use and MCU mockups, deterministic resets, and controlled credentials for repeatable agent work.
Capability evaluations
MCP-bench and TAU-bench extensions, TinyTAU for on-device agents, and task suites built around your real tools and constraints.
Trajectory data
Expert demonstrations, step-level annotations, preference labels, failure taxonomies, and calibrated evaluations for training and reward shaping.
Safety evaluation
Coding-agent safety, MCP injection assessment, computer-use injection red teaming, and adversarial trajectories with reproducible evidence.
Coding data
Repository-scale generation, issue resolution, code review, testing, debugging, tool-use, data analysis, and visual frontend tasks.
Built for
Agents across the whole tool chain
Task environments and data shaped around how each agent observes, reasons, uses tools, and recovers when the work stops going to plan.
Conversational agents
Natural-language dialogue grounded in evidence and policy.
Corporate assistants
Internal tools, workflows, knowledge bases, and governed automation.
Deep-research agents
Multi-source investigation, synthesis, and defensible conclusions.
Computer-use agents
Browsers, applications, filesystems, and realistic interfaces.
Coding copilots
Code writing, debugging, repository work, testing, and review.
OS & device agents
Desktop, mobile, embedded, wearable, and on-device environments.
How it works
A managed pipeline, built by engineers for engineers
01
Define
Objectives, constraints, rubrics, formats, and acceptance criteria become a testable specification.
02
Build & execute
We create the environment and tasks. Experts perform the work while the complete raw trajectory is recorded.
03
Validate
Automated checks enforce schemas, invariants, logical consistency, rubric adherence, and task completion.
04
Review
Senior reviewers audit difficult and flagged traces plus a statistically meaningful sample of the remainder.
05
Deliver
Versioned datasets, evaluation reports, deterministic environments, and audit logs arrive ready for your workflow.
Three ways to start
Data that meets you where the work is
Ready to deliver
Use an already produced and validated dataset. We align the format and delivery boundary with your stack.
Pipeline-ready
Start a prepared collection or evaluation pipeline, adjusted to your volume, domain, and acceptance criteria.
Expand or customize
Extend an existing corpus, adapt it to your domain, or commission a new environment and dataset from first principles.
Expert depth
Not crowd work. Technical work.
The work is produced and reviewed at the level where our own systems are built: Python, Rust, C/C++, JavaScript and TypeScript, Go, Java, Kotlin, SQL, Bash, mobile and ML stacks — across backend, frontend, systems, security, DevOps, data science, and model engineering.