Case study

PharmaBrain: Private, Air-Gapped LLM for Medical Research

Atacana needed a research assistant that could answer questions over thousands of proprietary news reports and drug discovery updates, without any data leaving its own servers.
Discuss a similar projectPrivate LLM deployments
  • Client: Atacana
  • Industry: Health & Fitness
  • AI capability: Generative AI
Visit Atacana (opens in a new tab)
Front page of a medical news publication, the kind of report PharmaBrain indexes for researchers
Offline, on local servers
100%
Parameter Mixtral MoE, 8-bit quantized
46.7B
Typical response time on local hardware
5–7s

The challenge

Thousands of new documents a day, and none of them could leave the building

Keeping up with medical research is a volume problem. Atacana's database grows every day with news reports, clinical trial updates and drug discovery announcements. Researchers needed to ask a question and get a grounded answer, not read hundreds of documents.

Standard cloud LLMs were ruled out by confidentiality requirements. The assistant had to run as a private, air-gapped deployment on local servers, and still be fast enough to be useful while thousands of documents were added daily.

  • Volume

    Thousands of reports and trial updates added to the database every day.

  • Data privacy

    Proprietary medical data could not be sent to a third-party cloud or API.

  • Speed on local hardware

    Answers had to come back in seconds without cloud-scale GPUs.

  • Grounded answers

    Every answer needed to come from the latest documents, not the model's memory.

What we built

A private RAG assistant on a quantized Mixtral model

PharmaBrain uses retrieval-augmented generation (RAG): it finds the relevant passages first, then has the language model answer from them. Every component runs on Atacana's own servers.

  1. Stage 1

    Index new documents locally

    News reports, trial updates and drug discovery announcements are ingested into a vector database on local servers as they arrive.
    Searchable document indexVector databaseLocal servers
  2. Stage 2

    Retrieve the relevant passages

    When a researcher asks a question, LangChain orchestrates the search and passes only the most relevant passages to the model.
    Context for the answerLangChain
  3. Stage 3

    Generate with a mixture-of-experts model

    Mixtral 8x7B has 46.7B parameters but uses about 12.9B per token, which makes it faster than dense models of similar size.
    Draft answer grounded in retrieved textMixtral 8x7B
  4. Stage 4

    Quantize for local hardware

    8-bit quantization cuts the memory footprint and processing time, so responses come back in 5–7 seconds on local servers.
    5–7 second responses8-bit quantization
  5. Stage 5

    Keep it air-gapped

    There are no calls to external APIs. Documents, questions and answers stay inside the client's infrastructure.
    Fully offline deployment
Research assistant · local server
Air-gapped

Incoming reports

  • trial-update-ax214.pdf
  • news-digest-0412.html
  • pipeline-review-q2.pdf
  • conference-abstracts.pdf
  • regulatory-notice-17.pdf

Vector index on local server

5 new documents chunked and embedded

Offline

1/4New reports are chunked, embedded and added to a vector index on the client's own servers. Nothing is sent to an outside service.

Illustrative sample: the documents and question are fictional. The model setup matches the deployment.

Results and impact

A static archive turned into an assistant researchers can question

PharmaBrain gives researchers fast, secure access to a large library of medical knowledge. It answers with sourced summaries of new developments in 5–7 seconds, fully offline.

  • Faster monitoring of drug advancements and trial updates
  • No proprietary data sent to a third-party cloud
  • Answers grounded in the latest documents in the database

It is a working example of the sovereign, private AI deployments we build for teams that cannot send data outside their network. Read how we handle data and IP on our security page.

  • Runs on local servers

    No dependency on public internet access or outside APIs.

  • 5–7 second answers

    Typical response time for a RAG query on local hardware.

  • Questions in plain language

    Researchers ask about acquisitions, trials and timelines the way they would ask a colleague.

  • Safe by design

    Retrieval keeps answers tied to real documents. See the guardrails we recommend for healthcare LLMs.
    LLM guardrails in healthcare

Project showcase

Sample questions and answers

Test questions from the project, with the model's answer next to the reference answer.

PharmaBrain sample questions, model responses and expected responses
#QuestionPharmaBrain responseExpected response
1When did Sobi acquire CTI BioPharma Corp.?Sobi acquired CTI BioPharma Corp. in June 2023.Sobi announced the CTI BioPharma acquisition on 10 May 2023 and completed it on 26 June 2023.
2What are the latest updates from the Ph3 PACIFICA trial?The completion date for the Phase 3 PACIFICA trial is expected to be by the end of 2026.Sobi mentioned that enrollment in the confirmatory Ph3 PACIFICA trial is projected to be completed by the end of 2026.
3When are the interim and final analyses of the Imetelstat Ph3 IMpactMF clinical trial expected?The Imetelstat Ph3 IMpactMF clinical trial is expected to have an interim analysis in 2024, and a final analysis in 2025.Interim analysis expected in H1 2025 and final analysis expected in H1 2026.
More healthcare AI workDocument AI for large archives

Keep exploring

Related case studies

All case studies

Book a strategy session

Talk to an AI engineer about your project

Tell us what you want to automate. The first call is a 30-minute working session with an engineer, not a sales pitch.

  • Send the form, it takes 2 minutes
  • We reply within 1 business day, under NDA if you need it
  • A 30-minute call to scope feasibility and next steps

Tell us about your project

By submitting, you agree to our Privacy Policy. We never share your details.