Home / Software & AI / Document Intelligence / Clinical Documents
Document Intelligence · Pharma manufacturing

We do the first pass on your batch records. Your reviewer decides.

The system checks every page of the batch record against the procedure — missing signatures, blank fields, out-of-range values, entries out of sequence — and hands your reviewer a marked-up pack with everything questionable already flagged.

A marked-up packevery question flagged with a page reference; your reviewer decides each one
5–6 weeksfrom first call to a specification your quality unit can test against
from $30,000Blueprint price, published — credited in full against the build
One document typewe start with a single batch record type, on your own documents

This is for you if

Qualified people are doing clerical checkingSignatures present, cells filled, numbers agreeing. Judgement work is not what is eating the day.
Cloud fails you on change control, not securityYou can answer the confidentiality question. You cannot answer “show me the system in use is the system you validated”.
Every change goes through procedureNothing deploys because someone merged it — and a supplier who learns that in month three is a problem.
You need to show where a value came fromThis value, this page, this document, this version, this reviewer.
Vendors keep proposing SaaSYou need a closed perimeter, and most of them have never deployed into one.

What the system checks, and what comes back

Three lists, because these are the three things a quality director asks in the first ten minutes.

What the system checks

Signatures and dates · required fields left blank · values outside specification · entries recorded out of sequence · disagreements between pages of the same pack · attachments that should be there and are not.

What your reviewer gets

A marked-up pack · a list of questions ordered by severity · a page and location reference for every question · everything else marked as checked. Anything the system could not read comes back as unreadable, human check required — never as a confident value.

What is recorded for the audit trail

The input · the model and rule versions in effect · what was flagged and on what basis · who decided what, on each question. Versions are pinned and dated from the first build, not assembled at handover.

The shape underneath: the model reads and proposes, deterministic rules verify, a person signs. The rules are the control — which is what keeps the validation package tractable, because a rule has a finite set of test cases and a model has a distribution. The full argument, and what validation demands of the architecture →

Where our responsibility ends and yours begins

This line is the whole commercial relationship. It is short on purpose.

OursYours
The system, the deployment, the pinned versionsValidation of the system — CSV and qualification
An update procedure that never reaches the networkAcceptance and the quality unit’s signature
The evidence chain, logged per recordThe decision on every flagged question
Batch release — always

How the work runs

Three phases, each with what we need from you. The first is a product you can buy on its own.

PHASE 01
Intended use and the check listWe write the intended-use statement with your process expert, take one batch record type, agree exactly which checks the system performs and which it must never attempt, and fix the acceptance criteria before anything is tested. Output is a build-ready specification, a fixed price, a timeline and the version-control approach your validation will reference. This phase is the AI Solution Blueprint, from $30,000, credited against the build; the spec is yours to implement with anyone.You provide
  • 30–50 completed records, the awkward ones included
  • The master records and SOPs they were made against
  • One process expert who can approve the intended use
PHASE 02
The checks, on your real recordsExtraction and the deterministic check layer built against your documents and measured on records you have already reviewed — with the miss rate and the queue volume reported as two separate numbers from the first week, because a single accuracy figure hides the one you are accountable for.You provide
  • A review owner who can adjudicate disagreements
  • Specification versions effective on the dates of manufacture
  • Retention and residency requirements
PHASE 03
Inside the perimeter, then handoverDeployment on your hardware behind the air gap, the import and version-inventory procedures written down, the exception queue integrated into your QMS or DMS, and the test-protocol inputs handed to your quality unit — which executes it and signs.You provide
  • Named owners for the version inventory and the import procedure
  • An acceptance owner who can sign
  • The hardware, or a quote from us for it

When we tell you not to do this

The first two disqualify more enquiries in this sector than everything else combined.

The records are still on paper and unindexedIf batch records live in binders and nobody has scanned them, the first project is scanning and indexing. That is a real project with a real supplier, and it is not us.
Nobody owns the specificationIf what the system is for has not been written down and approved by a process expert, acceptance testing cannot start — and that document, not the model, is what stalls these projects.
You want the decision automatedBatch release, disposition, any judgement on product quality. We will not build it, and a supplier who offers to should be asked how they intend to validate it.
The constraint is procedure, not throughputIf review is slow because the SOP requires three signatures in sequence, automation just makes a slow procedure faster to be slow at. Fix the procedure first.

How we deploy into a closed perimeter

Air-gapped is usually taken to mean “on our servers”. Properly it means no network path outside the perimeter — and everything the network used to do needs a replacement an inspector will accept.

What the network normally doesWhat we put in its place
Pulls model weights and updatesPhysical media transfer under a defined procedure — checksum verification, approval before import, an entry in the version inventory. No package manager reaches out.
Resolves dependencies at build timeA frozen, mirrored dependency set imported once and versioned. A build that cannot be reproduced offline is not a validated build.
Ships logs and telemetry outRetention inside the perimeter, with a controlled export path for the rare case where something must leave.
Provides time synchronisationA local time source. Audit-trail timestamps nobody can vouch for undermine every record they appear on.
Delivers security patchesA scheduled, approved cycle under the same import controls, with an agreed position on the lag between disclosure and application.

Two roles have to exist by name and usually do not at the start — an owner of the version inventory and an owner of the import procedure. We name both in the Blueprint. The hardware we supply directly: local inference nodes sized to the site’s document volume, quoted in the same contract as the software. See private AI infrastructure.

What we have already built

Client names withheld under NDA. See full case studies →

Life sciences · air-gapped compute

Zero-cloud AI for pharmaceutical R&D, in four weeks

A 12-person computational chemistry team whose IP counsel ruled out any cloud path — proprietary compound structures could not leave the perimeter, not even encrypted to a trusted provider. We supplied NVIDIA DGX Spark systems and RTX 6000 Ada workstations and took it from quote to a powered-on lab in four weeks. Read the case →

4 weeks quote to lab0 cloud dependency12 person team served
Aviation · MRO

A document pack checked against a regulation

The same mechanic under a different rulebook: pages classified into six document types, missing signatures and stamps detected, unfilled checklist cells found, numbers cross-checked across the pack — an annotated report, then a specialist signs. Advisory by construction, exactly as here. Read the case →

6 document types4 classes of checkCLI + API
Being precise about what these prove

The pharmaceutical project was a compute deployment, not a validated document system — it proves we put AI infrastructure inside a closed perimeter on a four-week timeline. The aviation project proves the document mechanic under a regulation, with the machine annotating and a specialist signing. Neither is a completed GxP validation package. A supplier who runs on someone else’s hosted endpoint cannot offer you either half; we would still rather you knew the limits of ours before the first call than after it.

Frequently asked questions

Have you been through a CSV validation, and what exactly do you hand to our quality unit?

We have not completed a GxP validation package as a supplier, and we would rather say that here than have you find out on a capability call. What we hand over is the material that makes your validation tractable: an intended-use statement approved by your process expert, a written specification with acceptance criteria agreed before testing begins, pinned and dated versions of the model, prompts, rules and runtime, a build that reproduces offline, a deterministic control layer that is cheap to qualify, and an evidence chain per record. Your quality unit executes the protocol and signs. Responsibility for validation evidence rests with the regulated company regardless of who built the system, so this is the correct division rather than a limitation of ours.

How is a project in a validated environment scoped and priced?

Scope first, price second, and both numbers are published. The specification is a product: the AI Solution Blueprint, one system, 5–6 weeks, from $30,000, credited in full against the build if implementation starts within 90 days. In a validated environment it also produces the intended-use statement, the version-control approach and the acceptance criteria your quality unit will test against. Builds of this kind typically start around $150,000 over 4–6 months, plus hardware where the deployment is air-gapped.

Why can we not just use a cloud AI service?

Usually the blocker is change control rather than security. A hosted model is updated on the provider's schedule: a system qualified against one version can be running on another the following week, with no notification and no record of when it changed. That is an uncontrolled change to a qualified system whatever the new version's quality. Add the inability to freeze the surrounding service and the retirement of versions you qualified against, and the evidence chain has holes you cannot close from your side.

What does air-gapped deployment involve on your side?

Local inference hardware, which we can quote in the same contract, plus the procedures that replace everything the network would have done: physical media transfer with checksums and documented approval, a frozen mirrored dependency set so the build is reproducible offline, retention of logs inside the perimeter, a local time source, and a named owner for the version inventory. The hardware is the straightforward half; the procedures are what an inspector asks about.

Will the system decide anything about product quality?

No, and we will not build it that way. Batch release, disposition and any judgement on product quality stay with the qualified person. The system reads, checks against a specification and presents what a human must look at — it decides what needs attention, never what the answer is. If a supplier offers to automate the decision itself, that is the point to ask how they intend to validate it.

How much of the work do rules do rather than a model?

More than in any other document process we build, for a specific reason: you have to validate whatever acts as the control, and a rule is far cheaper to qualify than a model. A rule has a specification and a finite set of test cases; a model has a distribution. So the model reads unstructured input and proposes, deterministic rules verify, and the rules are what your validation package covers.

Related practices

Start with one document type

Pick one batch record type. We map it, build the check against your procedure and show you the marked-up output on your own documents — inside the Blueprint, from $30,000, credited against the build.

Not sure a project is the right first step? Free AI Readiness Score — 3 minutes.

An engineer replies within one business day. Prefer email? sales@haink.org