ACCESS IS TOO BROAD
People get wide access because narrowing it is hard, and the platform was never built to make it easy.
USE CASE / HEALTHCARE & LIFE SCIENCES
Sensitive data, strict access, and reporting that has to be right the first time.
In healthcare the cost of a wrong number is not a bad quarter. Every design decision has to assume the data is sensitive, the access has to be provable, and somebody will one day ask you to reproduce a report exactly as it looked two years ago.
What usually goes wrong
IDENTIFIERS HIDE IN FREE TEXT
Names and identifiers appear inside notes and comment fields, so masking the obvious columns is not enough.
REPORTS CANNOT BE REPRODUCED
Rerunning last year's report gives a different answer, because the underlying data has been updated since.
CONSENT IS TRACKED BY HAND
Whether a record may be used for a given purpose is recorded in a spreadsheet, or in somebody's memory.
AI STOPS AT THE LEGAL REVIEW
Every AI idea dies in review, because nobody can prove where the data would go or who would be able to see the output.
What we build
- De-identification applied at the moment data arrives, before it ever lands in the analytics layer, not as a step someone remembers later.
- Role-based access built into the platform, so permissions are a property of the data and not of each individual report.
- A full, unchangeable audit trail of who read what and when, produced automatically.
- Point-in-time reporting, so any report can be reproduced exactly as it looked on any past date.
- Consent and purpose recorded as data, and enforced by the platform, so a record cannot be used for a purpose it was not permitted for.
- A documented data flow diagram that a legal and compliance reviewer can actually read, produced as part of the build rather than after it.
What AI looks like here
Protocol and SOP assistant
- What it does
- Lets clinical and operations staff ask questions across protocols, standard procedures and guidelines, and always shows the exact source paragraph next to the answer.
- How it works
- Approved documents are split into sections and stored as embeddings in a vector database inside your own cloud account. A question retrieves the relevant sections and a model writes the answer from those sections only, with the source shown. Nothing is answered from the model's general knowledge.
- What it needs
- A controlled document library with clear versioning, a model hosted inside your own network boundary, permission rules per document, and a logged record of every question asked.
Identifier detection in free text
- What it does
- Scans free-text fields and flags where names, contact details, identifiers or dates of birth are hiding, before that data reaches the analytics layer.
- How it works
- A detection model reads each free-text field and marks the spans that look like identifiers. Anything flagged is masked or redacted automatically, and a sample is routed to a human reviewer each week to check the model is still accurate.
- What it needs
- A detection model that runs inside your own environment, a masking step in the pipeline that runs before the data is stored, and a regular human review of what it caught and what it missed.
Operations exception summary
- What it does
- A daily plain-English summary of what needs attention: which sites are behind, which data-quality checks failed, and what changed since yesterday.
- How it works
- The platform produces the exception list as structured data. A model turns that list into short readable paragraphs following a fixed template. No patient-level data is included, only counts and site-level figures.
- What it needs
- Reliable daily loads, agreed thresholds for what counts as an exception, and a hard rule that no individual-level data is ever included in the generated text.
The tools we use here
- Databricks with Unity Catalog, or Snowflake with dynamic masking
- AWS Glue or Azure Data Factory
- dbt
- Apache Airflow
- Terraform
- Amazon Bedrock or Azure OpenAI, both hosted inside your own network
- Private endpoints
- Customer-managed encryption keys
- Full audit logging
What changes for you
- Access reviews take an afternoon instead of a fortnight, because the platform can answer who saw what.
- Sensitive fields are protected by default, not by somebody remembering.
- A report from two years ago can be reproduced exactly, on demand.
- AI projects get through legal review, because the data flow was designed to be reviewable.
What we will not do
- We do not send patient or subject data to any model outside your own cloud account.
- We do not build anything that produces clinical advice or influences a care decision.
- We do not rely on a model alone for de-identification. Detection is checked by people.
- We do not start an AI build before the access and consent model is agreed in writing.