Ask anything about your study’s data.
An agentic AI analyst with the whole study in scope — every scale, every rater, every visit, every transcript — and clinical-trial tools built for it. Not a fixed set of dashboards, and not a general-purpose chatbot: it computes each answer from the real data, inside an environment you control.
Request a demoWhat you can ask
Every dataset in the study is in scope.
Scores and model predictions, item-level discordance, questionnaire adherence, protocol visit windows, interview transcripts, secondary endpoints, participant disposition — and wearable streams where a study collects them. One place to ask, whatever the question is about.
Built for this data, not pointed at it.
A general-purpose chatbot can read a file you paste into it. Mira has typed, validated access to the study's own tables — the same ones the platform renders — through tools written for clinical-trial work: per-site metrics, model–rater discordance, adherence, visit windows, transcript lookup.
And when no tool fits, it writes one.
Where no view in the platform precomputes the answer, Mira writes the analysis itself — a rolling average, a custom cohort, a distribution nobody had thought to build — and runs it over the full dataset in an isolated sandbox.
How it answers
It shows its work while it works.
A real question takes real analysis, so Mira narrates each step as it runs — which data it opened, what it computed, how long it took. You see the path to the number before you see the number, and every site, rater and participant it names is a link straight into the data.
Four sites sit in the top quartile for enrolment rate and the bottom quartile for scoring agreement. Site 07 is the clearest case — it is enrolling 2.4 participants a month against a study median of 1.1, with a mean difference of 3.4 points.
| Site | Participants | Enrol / month | Mean diff | Visits |
|---|---|---|---|---|
| Site 07 | 19 | 2.4 | 3.4 | 112 |
| Site 03 | 16 | 2.1 | 2.6 | 94 |
| Site 19 | 14 | 1.9 | 2.3 | 81 |
| Site 11 | 13 | 1.8 | 2.2 | 77 |
It sees what you are looking at.
Open Mira beside any page and it already has your view: the filters you set, the rows on screen, the participant visit in front of you. Ask “why is this one flagged?” and it answers about that visit — no re-specifying, no copying IDs across. The nearest thing to an analyst reading the same screen over your shoulder.
Answers you can take away.
Charts and tables render natively, not as text you have to reformat. Any result exports to CSV, and every entity it names links back into the platform so the next person can check the work.
It also looks on its own
For the questions you didn’t think to ask.
Knowing what to ask is its own problem. So overnight, Mira works through the study the way a reviewer would and leaves a short briefing — a few findings worth a look, each already analysed. Click one and the conversation opens with the reasoning in it, ready for your follow-up.
Model–rater difference rose to 3.4 points this week
Up from 2.1, concentrated in two raters.
Scales disagree on 18% of visits at four sites
Against a 6% study average.
Questionnaire completion recovered to 94%
Back from 78% last month.
Why the numbers hold
It computes. It does not recall.
A language model doing arithmetic in its head is a liability in a clinical tool. Mira is built so that it cannot.
Every figure is computed server-side.
Counts, averages, trends and distributions are calculated against the study data by the same services that render the dashboard — never estimated by the model, and never summed from a sample of rows.
It cannot reach past what you can see.
Every tool is typed and validated. An unknown column is rejected rather than quietly swapped for a near match, and the agent can never read a rawer source than the platform already shows you.
It says when it has nothing.
If a query returns nothing, Mira says so plainly instead of filling the gap. Text inside the data is treated as data — a rater comment that reads like an instruction is never followed.
Security
Asking anything, without opening anything up.
The reason a general-purpose assistant cannot do this job is that it would need your study data to leave your environment. Mira’s does not. It is deployed inside your own cloud or our dedicated Virtual Private Cloud, scoped to one study at a time, and bounded by the same access rules as the rest of the platform.
The riskiest part — running analysis code the model wrote — is deliberately given nothing worth stealing.
Deployed in your environment
Your own cloud or our dedicated Virtual Private Cloud. Study data stays inside a compliant environment you control, and the model is called within that deployment’s own cloud project and region.
Scoped to one study, and to you
Each study is served by its own isolated service. Every request carries your identity and is checked against your study access before any data is read — the agent inherits exactly your permissions, never more.
Analysis runs with nothing to steal
Model-written code executes in an isolated container with no credentials, no database or storage access, and all outbound network denied. It receives only the data for that one question and hands back only the result.
Auditable end to end
Every question, every tool call and every access is recorded. Nothing the agent did is a black box after the fact — including which rows an answer was computed from.
Bring us your hardest question.
We run Mira against your blinded data and you ask it whatever you would normally wait a week for.
Request a demo