AI for regulated industries
Meet CairaCaira is the assistant inside your quality and laboratory platform. In a GLP test facility, a GMP plant or a pharmaceutical testing lab, she reads only the records you allow, never sees a name, and writes every answer into a tamper-evident audit trail.
Caira · GLP test facility
Sample data. Names, studies and record numbers are fictional.
Why regulated labs can switch her on
A model can be told to behave. Caira is built so that it cannot misbehave with your data: the limits are checked on every call, before any record is read.
Caira reads only the objects and fields an administrator enabled, intersected with what the user may already see in the platform. A field that isn't listed is never fetched, whatever the model asks for. Users without "view all" see only records they own.
Names, emails, phone numbers and IDs become stable tokens before anything leaves your server. The model reasons about [EMPLOYEE_2]; you read "Meera Pillai". The map between the two never leaves the server.
Every turn is one row: who asked, what was read, which document versions were in force, which model answered, and the reply. Each row is hashed together with the one before it, so Verify chain finds the first row that was changed or removed.
Caira has three read-only tools: count records, list recent records, query records. There is no tool that writes, so nothing she says can change a record.
Where she works
Caira reads what your platform records. The policy, the masking and the audit trail are the same everywhere; only the record types differ, and you decide which ones she may see.
OECD GLP · 21 CFR Part 58
Studies, study plans and amendments, dose groups, in-life data, deviations, CAPAs, QA inspections and the master schedule. A management page shows the whole facility at once.
"Which studies are late to start?"
EU GMP · 21 CFR 210/211
Deviations and investigations, CAPAs and their effectiveness checks, change requests, complaints, internal audits, calibration and environmental monitoring.
"Which change requests are waiting on QA?"
ISO/IEC 17025 · GMP QC
Sample requests and samples, test methods and specifications, results and witnesses, instruments and their intermediate checks, logbooks and analyst skills.
"Which samples are past their turnaround time?"
And more
Calibration laboratories, controlled documents and training records, R&D project records: wherever entries are signed under 21 CFR Part 11 and have to be explained later.
"Who signed the last version of SOP-014?"
How it works
Eight checks between a person typing and an answer appearing. The example follows a GLP facility; a GMP site or a QC lab goes through the same eight. Press a step, or let it play.
What she does for a facility
Counts, lists and lookups across deviations, CAPAs, change requests, studies, samples, instruments and schedules, in plain language. She says which date field she used and reports the number the tool returned, exactly.
Instructions, refusal text and knowledge documents go through draft, review and approval, like an SOP. Only approved versions whose effective date has passed reach the model, and each turn records the version and hash of every document in force.
Administrators write deterministic checks: a value outside its reference range, one person booked twice. Rules raise findings; a named person accepts or rejects each one, with a signature where required. The model never decides what is flagged. Try one
For GLP test facilities today: one page shows what is overdue, how loaded the schedule is and which studies are at risk. Caira writes a short brief from the page's figures and checks every number before it is shown. See the page
A behaviour eval set run through the live pipeline, unit tests for the deterministic parts, a verifiable audit chain, and usage metering with an allowance alert. See the chain
Whether a body-weight drop is treatment-related is the Study Director's call. Caira refuses that kind of question before the main model runs, and the refusal is logged. She reads, explains and flags; your people decide.
Test Facility Management · GLP
Built for GLP test facilities, where management carries overall responsibility for compliance. The figures are counted by fixed rules, so they are the same every time and can be explained to an inspector. Caira's job is what counting can't do: read the whole picture and make it easy to act on.
Caira's brief
Every number and name below was checked against the page. Hover one to see where it comes from.
| Type | Record | Study | Owner | Due | Days |
|---|
Also on the page: schedule load per person and per study, studies at risk with a "Why?" that Caira explains from the record, draft reminders to owners, suggested reassignments with one-click apply and undo, week-on-week trends, and a printable meeting pack. Ask us about management views for GMP and QC sites.
Rules and findings
A rule is a fixed, tested check written and approved by an administrator. Move the value below outside its reference range and watch what Caira does, and what she leaves to you.
MeasurementRange)Within range. Nothing to flag.
Open → Accepted or Rejected → Cleared when the next run no longer finds it
Findings are never deleted and never duplicated: the same record and rule raise one finding until it is cleared. Deciding one needs edit rights on the record, and a password signature where the rule requires it. Each decision is appended to the audit chain.
The audit trail
The hashes below are computed in your browser right now, with SHA-256, exactly as the platform computes them. Change the question in a row, then press Verify chain.
| # | Who | Question | Tools · outcome | Previous | Hash |
|---|
Edit the question in row 3, then verify.
Who asked, the question as the model saw it, the scope label and reason, the capability, every tool call with its arguments and the record ids it returned, calls the policy denied, the hash of the prompt and the version and hash of each document in force, the provider and model, tokens in and out, the reply, and the outcome.
The chain is tamper-evident, not tamper-proof: it shows that a row was changed or removed; your database controls prevent it. Tokens on every row feed usage metering, with one alert mail at 80% of the monthly allowance.
Inside your platform
Caira is a product of Fiscal Ox and runs inside the Fiscal Ox quality and laboratory platform, next to the records she reads. There is no copy of your data in a separate AI system.
Questions we are asked
Any site whose records are kept for an inspector: GLP test facilities under OECD GLP and 21 CFR Part 58, GMP manufacturing under EU GMP and 21 CFR 210/211, pharmaceutical QC and contract testing laboratories under ISO/IEC 17025, calibration laboratories, and the quality systems around them (deviations, CAPAs, change control, complaints, audits, controlled documents). Caira reads the record types your platform holds; an administrator decides which ones she may see, field by field.
Only what is in the request to the model: the scrubbed question and the stored transcript in its token form, the prompt (platform rules, the capability list, refusal text, approved documents, today's date, the objects and fields this user may read, the user's role and profile, and the user's own name as a token), and tool results after masking. The username is never sent, and the token map stays on the server.
The one you choose on the integration page, called with your own key. Today that is Groq (default) or Anthropic; the chat loop, masking, policy and audit are the same whichever is in use. The model name is recorded on every audit row, so a change is visible in the trail.
No. Her three tools count, list and query. There is no tool that writes. On the facility page, a suggested reassignment is applied only when a person presses Apply, through the platform's normal permission check and record audit trail, with undo.
Three layers. The data layer makes a denied field unreadable whatever the model says. The scope check refuses scientific-judgement and out-of-scope questions before the model runs. And on the facility page, every number, record and person in Caira's brief is checked against the page's own figures; a brief that fails the check twice is not shown at all.
The deterministic parts (policy, masking, chain, scope-check parsing, scoring) have unit tests. The model-dependent part has a behaviour eval set: greetings, in-scope questions that must use a tool, out-of-scope and scientific-judgement questions that must be refused, an impersonation attempt, a PII probe and injection attempts. You run it from the setup page as a signed-in admin after any change to the model, scope, a capability or a document, and keep the logged result.
It sets no cookies, loads nothing from third parties, and runs no analytics. The fonts are served from this site. The demos on this page run entirely in your browser with fictional sample data.
Request a demo
Forty minutes, on your platform or ours: the chat on your record types, a management page, a rule raising a finding, and the audit chain verified in front of you.
Or write to hello@caira.si.