Frema Labs

By industry

By technology

By company size

Start from your industry, the technology you are weighing, or the size of your company.All solutions
Let's talk

Solutions / Pharma and life sciences

AI for Swiss pharma and life sciences teams

Tools that fit validated processes, with the documentation your quality unit will ask for.

Describe your case

What we could build

These scenarios are illustrations. None of them describes a client.

Pharmacovigilance

Adverse event reports arrive as emails, forms and literature excerpts, and case processors retype each one into the safety database.

What we would build

An intake pipeline that extracts reporter, patient, product, event and seriousness from each report, proposes MedDRA terms and highlights the source sentence for every field. Processors review the pre-filled case before it enters the safety database through E2B(R3) import, and the pipeline is validated for that intended use.

Mistral, MedDRA, Python

Regulatory affairs

Answering a health authority question means searching years of submissions, SOPs and guidelines for the one paragraph that settles it.

What we would build

A search assistant over your submissions, SOPs and Swissmedic, EMA and ICH guidelines that answers in a few sentences and links each claim to the exact section. It is read-only, and it declines to answer when the documents do not support one.

FAISS, bge-reranker-v2-m3, vLLM, Veeva Vault

Quality assurance

QA reviewers page through executed batch records line by line, looking for a missing signature or an out-of-range value.

What we would build

A review agent that reads scanned batch records, compares each step with the master record and lists deviations, missing signatures and out-of-range entries with their page. The reviewer works through that list, and release stays with the person who certifies batches.

Qwen3-VL, LangGraph, Python

Manufacturing science

A bioreactor run that starts to drift is usually noticed once the batch is already off track.

What we would build

An anomaly detection model on process time series that learns the normal course of a run and warns when a signal starts to diverge, naming the sensors involved. Engineers see which signals moved and when, and decide whether to intervene.

PyTorch, Darts, AVEVA PI, MLflow

Medical affairs

Scientists are meant to follow every new paper in their field, and the weekly volume makes that impossible.

What we would build

A literature agent that queries PubMed every week for each team's topics, screens abstracts against their criteria and sends a short digest with links and the reason each paper was kept. Scientists rate the picks, which tunes the next week's screening.

PubMed API, BGE-M3, LangGraph, Microsoft Teams

What we design around

In a regulated process, a model is only useful if it can be validated, versioned and explained to an inspector. We design for that from the first week: a narrow intended use, a test set your experts agree with, a fixed model version, and a person who signs every output that matters.

  1. GxP validation

    Systems that touch a regulated process need a defined intended use and a validation plan, so we write those documents with the software and follow your change control.

  2. Swissmedic and EMA

    Answers and documents often have to cite the exact guideline or submission they rely on, so every output carries its sources.

  3. Patient data

    Safety reports and clinical data are sensitive personal data under the nFADP, so processing stays in your own environment or with a host your data protection officer approves.

  4. Human sign-off

    Case processors, QA reviewers and scientists stay accountable, so every scenario ends with a person reviewing the result.

Questions we hear

Yes, if the intended use is narrow and the model version and settings are fixed. Following GAMP 5, we test against a set your experts approve, document the limits, and rerun the tests on every change.

Yes, within your environment or with Swiss hosting. We define the data flows with your data protection officer before any data moves.

The assistant is instructed to answer from retrieved passages only and shows them next to the answer, and a check flags any sentence they do not support. When nothing relevant is found, it says so.

Usually with literature monitoring or document search. Both need little integration and help a small group of scientists from the first weeks. Our SME page, linked below, covers team and budget at that size.

Related solutions

Technologies

Industries

Company size

Which document does your team spend the most time searching for?

We start with a feasibility study on your own data, so you see results before anything is promised.

Describe your caseHow we work