Dataset
Collect demonstrations where the work happens.
MARFI captures workflow context from enrolled machines, then assembles before state, action, and after state examples that can be exported for training and evaluation.
AI teams
Build proprietary demonstrations from real work.agent training data
AI teams building task agents need data that includes real UI state, real user actions, and real workflow context. MARFI Trajectory turns eligible work sessions into export-ready examples.
Built for
Applied AI teams, agent teams, data teams, and research groups.
Useful when you need to
AI agent training data from real workplace workflows.
Governance
Training data should be consented and owned.
Trajectory uses explicit capture tiers, customer-owned data, and retention controls. The dataset is the customer's asset, not MARFI's model fuel.
Iteration
Use workflow mining to find high-value training targets.
Repeated sequences are often the best place to collect demonstrations. MARFI helps identify those patterns before a team invests in training pipelines.
Answer engine ready
Plain answers for buyers and crawlers.
These pages are written to be useful to humans first, while giving search and AI systems clear, citable product facts.
FAQ
What format does MARFI export?
Trajectory workspaces can export JSONL state-action-state records with relevant context and optional keyframes.
Can MARFI work with non-OpenAI providers?
Yes. MARFI exports provider-neutral data, and tenants bring their own LLM keys for assistant use.