use cases · Anonymisation
A dataset you cannot share is a dataset you cannot use twice. Detect the personal data in documents, transcripts and records, replace it consistently, and keep the text useful for the analysis you wanted to run on it.
how it works
Three steps, with the component that does the work in each one.
step 01
deepseek-v4-flash A regular expression finds a phone number. It does not find "the patient from the Tuesday clinic whose brother works here", which identifies a person just as well. Reading the sentence is the part that needs a model.
step 02
pseudonym map The same person becomes the same token everywhere in the document, so the text still makes sense and the analysis still works. Consistency is what separates a usable pseudonymised file from a page full of black boxes.
step 03
qwen-3.6 A second pass over the redacted text looks for what the first one left. This step is not optional: the cost of a miss here is a notification to a supervisory authority, and a sampling review does not find a miss in document four hundred.
drop-in
Point the OpenAI SDK, or LangChain, LlamaIndex, your own pipeline, at Helmcode. Same calls, same shapes, private models on EU infrastructure.
read_the_docsfrom openai import OpenAI client = OpenAI( api_key="sk-...", base_url="https://api.helmcode.com/v1", # one line changes ) # 1 · find what identifies a person, not just a pattern found = client.chat.completions.create( model="deepseek-v4-flash", messages=[ {"role": "system", "content": FIND_IDENTIFIERS}, {"role": "user", "content": document}, ], response_format={"type": "json_object"}, ) # 2 · same person, same token, everywhere in the file clean = apply_pseudonyms(document, found) # 3 · a second pass over the redacted text check = client.chat.completions.create( model="qwen-3.6", messages=[{"role": "user", "content": clean}], )
why here
What changes when you run it on open models, on infrastructure you control.
Anonymisation is the one job whose input is by definition the unprotected version. Sending that to a third-party API to have it cleaned is the problem you were trying to solve, which is why this case usually runs on-premise.
Nobody anonymises ten documents, they anonymise an archive. A flat rate means the size of the archive stops being the reason the project never starts.
Redaction reduces risk, it does not automatically make data anonymous in the legal sense: that is an assessment about re-identification risk, and it belongs to your data protection officer. What infrastructure gives you is a process you can document and audit.
industries
The sectors where this case has traction, each with its own page and its own regulation.
// faq
What data protection and clinical teams ask before letting a model touch the raw version.
That is what consistent pseudonyms are for. If every mention of one person maps to the same token, you can still count, group and follow a case through a file. Blanket masking destroys exactly the structure you needed.
A multimodal model reads the page image directly, so a scan does not need a separate OCR stage before it can be processed. Handwriting is the hard case and it is worth measuring on your own material before committing.
No. Prompts and outputs are never stored and never used for training. On the on-premise deployment the question does not arise, because nothing leaves your network in the first place.
// get started
Skip the AI infra work. Deploy your first private inference endpoint today.
Flat rate. EU data. OpenAI API compatible.
// cookies
We use strictly necessary cookies to run the site and, only with your consent, Google Analytics to understand usage. No advertising, ever — see our Cookie Policy.
// preferences