Training data from the whole company, not one tool.
De-identified datasets from established businesses across industries. Each one is linked across email, documents, tickets, projects and accounting records, and licensed in writing.
Whatever you're training, we'll source it.
We work across industries and data types.
Every dataset is a whole company.
Each dataset follows one business across the tools it runs on, for years, so a model sees the full path of the work and how it ended.
- Systems
- Email and chat, documents and files, support desks, projects and operations, CRM, and billing and accounting.
- Tools
- Google Workspace, Microsoft 365, Slack, Zendesk, HubSpot, Salesforce, QuickBooks, Procore, Jira and others.
- History
- Years of continuous records, not snapshots.
- Linking
- Records linked across systems, so one piece of work can be followed from the first message to the final invoice.
- Outcomes
- The result stays attached: the quote that became an invoice, the ticket that got resolved, the bid that was won or lost.
- Source
- The private systems of operating businesses, so it's clean for evaluation as well as training.
We're selective about what we take.
A dataset is only as good as the company behind it. We look for established businesses with long, continuous histories across several systems, and we turn down records that are thin, fragmented or too sensitive.
What we look for
- Years of continuous records
- Several connected systems
- Enough day-to-day volume to learn from
- Clear rights to license
- Thin history
- Fragmented systems
- Unclear rights
- Too sensitive
Pseudonyms that hold across systems.
People and companies get realistic stand-ins that stay the same across email, tickets and invoices, and across years. Conversations still read naturally, and the links between records survive.
Licensed in writing.
License- Licensed at the source
- Every dataset is licensed in writing from the company it comes from, and that company confirms in writing that it has the right to license it.
- De-identified only
- Raw records never go to AI teams.
- Limited use
- Training and evaluation, on terms set in writing for each dataset.
- No re-identification
- Every license prohibits attempts to identify the people or companies behind the data.