AI that makes the evidence cleaner, never less defensible.
Comphas exists so every number can be defended: what was counted, on what date, from which source. AI that silently edits data or invents a finding would undo that. So AI in Comphas follows one rule.
AI proposes. The analyst decides. Comphas records.
Where AI saves the most time at the lowest risk.
Design assist, clean and enrich, and interpret. Each one works inside the same rule, so the evidence stays defensible.
From a question to the right chart
Ask in plain language. Comphas proposes the filters, the use case and the chart type that fits the data.
- "Who is recruiting Crohn's patients in Germany?" becomes a recipe
- Uses the 91 chart types' own rules
- Refuses a chart the data can't support
- Changes no data
Less noise, fully traceable
Sponsor aliases, free-text conditions and intervention names matched against reference sources.
- Matches against MeSH, RxNorm, ChEMBL and Open Targets
- Each change is a proposal with a confidence score
- No match means a flag, never a guess
- Accepted changes become rules that replay, with the raw record kept
A draft reading of every chart
A finding title and a short read, written for the analyst to edit and approve.
- Every number comes from the query, never from the model
- No causal claims; calibrated wording only
- Marked as a draft until an analyst signs off
- Only approved text can be frozen into a pack
Proposed, reviewed, then replayed.
The analyst sees every proposal with its source and confidence. They accept, reject, or approve everything above a threshold. Accepted proposals become a rule set that runs the same way on every refresh.
The rule set is part of the method record and the audit trail. A frozen pack names the rule set that produced it.
Illustrative example of the workflow with sample data.
Where the model runs matters.
Questions about an acquisition, a safety signal or a programme stop are exactly the questions that must not leave the building. That decides the architecture before any feature is built.
- Inside your instance. Models run within the client's single-tenant environment, not through a shared public chat service.
- Your data stays yours. Client data and questions are not used to train models.
- Nothing silent. No AI output changes a dataset, a chart or a pack without an analyst's approval.
- Always traceable. Every accepted proposal is logged with its source, confidence and approver.
Shape it with us.
Pilot customers can help set the thresholds and reference sources that matter in their work.