Regulated AI & Compliance

Regulated AI and medical device software built for FDA and MDR review

We audit the code and documents behind your software as a medical device (SaMD) or GxP system, rework what a reviewer would reject, and write the IEC 62304 lifecycle files, validation and risk evidence your submission needs. Then we stay until the certificate arrives.

  • 5 stages from code audit to certificate
  • +21% diagnostic accuracy with our breast imaging AI
  • 32% fewer readmissions within 90 days of a pilot
  • 5.0 Clutch rating, 22 client reviews

Who we work with

Four teams that answer to a regulator, each with a different deadline

  • 01

    SaMD startup

    Your MVP works, but a notified body would reject the repo. We make it reviewable without a rewrite.

    Audit my codebase
  • 02

    Medtech manufacturer

    AI is joining an existing device. We extend your risk file and software lifecycle to cover the model.

    Add AI to a device
  • 03

    Pharma and biotech

    Models touch GxP data or batch records. We validate them under GAMP 5 and Part 11 rules.

    Validate GxP AI
  • 04

    Clinical software vendor

    Hospitals ask whether your tool is a device. We map the claim and the evidence behind it.

    Check device status

At a glance

Medical device software development in brief

  • Who it is for

    Teams whose software diagnoses, treats, monitors or controls GxP data, and must show a regulator how it was built.

  • What you get

    Code and files a reviewer can trace, from each requirement to its test, risk control and release.

  • Where to begin

    A fixed-scope audit of one product: code, architecture, documents and intended-use claims.

  • What it costs

    Audit from €5,000, a fixed fee for 2–3 weeks. Rework, submission support and the retainer are quoted from the audit report.

Quick check

Your product needs this if… and when to wait

You need this if

  • Its output changes a diagnosis, therapy or dose.
  • A notified body, FDA reviewer or GxP inspector has asked for evidence you cannot produce.
  • The model was trained before anyone wrote down requirements or a risk analysis.
  • You plan to retrain after launch and need a change plan regulators accept.

Skip it if

  • Your software is wellness or admin only. AI Transformation fits non-device workflows.
  • You need a regulatory strategy only. A regulatory consultancy can file it; we fix the engineering behind it.
  • You need company-wide AI policy. AI Governance covers all the AI your company uses.
  • You want an off-the-shelf QMS. We fill your eQMS with evidence; we do not sell one.

Delivery path

From code audit to certificate: five stages, one team

Audit, Rework, Document, Submit, Sustain. The audit is a fixed-scope project. Each later stage is quoted from what it finds, and you can stop after any of them.

  1. 01

    Audit

    2–3 weeks · fixed fee

    We read the code, architecture and existing files against IEC 62304, ISO 14971 and your intended use.

    Deliverable: a gap list with rework scope and effort.

  2. 02

    Rework

    Scoped by the audit

    We refactor, add tests and traceability, and fix what a reviewer would flag first.

    Deliverable: code that matches its own documentation.

  3. 03

    Document

    In parallel with rework

    We write the lifecycle, risk, validation and cybersecurity files with your QA team.

    Deliverable: a design history file ready for review.

  4. 04

    Submit

    With your RA lead

    We answer the technical questions raised on a 510(k), De Novo or MDR technical documentation review.

    Deliverable: engineering input for each reviewer question.

  5. 05

    Sustain

    Monthly retainer

    We keep the model, change plan and post-market evidence current until and after certification.

    Deliverable: releases that stay inside the approved plan.

Design history file

Which files will the reviewer open? The ones we write with you

We write each file from the code as it is, so the evidence matches the release you ship. Templates follow your QMS, not ours.

Software lifecycle plan
IEC 62304 activities, safety class and SOUP list.
Risk management file
ISO 14971 hazards, including model failure modes.
Traceability matrix
Requirement to design, test and risk control.
Model validation report
Datasets, metrics, subgroups and limits of use.
Change control plan
How the model may change after release.

Need controls across all your AI, not one product? AI Governance covers the whole estate.

What’s included

Regulatory scope

Pick the parts your product needs. Most teams start with the audit and the lifecycle files.

  • SaMD and SiMD IEC 62304, IEC 82304-1, ISO 14971.
  • FDA pathways 510(k), De Novo and PCCP inputs.
  • EU MDR and IVDR Technical documentation for notified bodies.
  • GxP and Part 11 GAMP 5 validation, audit trails, e-signatures.
  • Cybersecurity SBOM, threat model, premarket security files.

How we price

What does compliant medical device software cost?

The fit call is free. The audit has a fixed scope and a fixed fee from €5,000, agreed before you sign. Rework and documentation are quoted from its gap list. The retainer covers the months between submission and certificate.

  • Fit call

    €0

    30 minutes to check device status, risk class and which audit scope fits.

    Bring your intended use

  • Code and file audit

    From €5,000, fixed fee

    Code, architecture and documents reviewed against IEC 62304, ISO 14971 and your pathway.

    2–3 weeks · gap list you keep

  • Certification retainer

    Quoted from the audit

    Reviewer questions, model updates and evidence upkeep until and after the certificate.

    Monthly · terms agreed after the audit

Every audit includes: a gap list ranked by reviewer risk, a rework estimate and a call with your QA and RA leads.

Rework and documents: quoted per work package, so you can split them between us and your team.

What sets the fee: the size of the codebase, the software safety class and how many documents already exist. We fix it in writing after the fit call.

Confidentiality: we sign an NDA before the audit, and a data processing agreement, or a HIPAA business associate agreement, before we handle personal data or PHI.

Already have a regulatory consultant? We work under their strategy and supply the engineering evidence they need.

Prices exclude VAT where it applies. We are an engineering team, not a notified body or law firm; clearances and certificates come from FDA and notified bodies.

Regulated AI or AI Governance? One product, or your whole AI estate.

Regulated AI and Compliance compared with AI Governance Consulting
CompareRegulated AIAI Governance
ScopeOne device, SaMD or GxP systemEvery AI system in the company
RegulatorFDA, notified body, GxP inspectorInternal audit, buyers, EU AI Act
DeliverableDesign history file and reworked codeRisk register, policies and controls
Price modelAudit from €5,000, fixed feeGap assessment from €5,000, fixed fee

Clinical and pharma proof

AI we have shipped where errors reach patients and labs

Each project below ran on clinical or laboratory data, where a wrong output has a cost beyond the software. Figures are taken from the published case studies.

  • +21% diagnostic accuracy, radiologist with AI vs. radiologist alone
    Healthcare · medical imaging, Canada

    Breast cancer detection from mammography and ultrasound

    Detection accuracy rose 20–30% in clinical trials when radiologists used the tool. See the case: breast cancer detection platform

  • 32% fewer readmissions within 90 days of the pilot rollout
    ICS · integrated care platform

    Readmission prediction built into a clinical platform

    84% of patients flagged as high risk were in fact readmitted, so care teams could target follow-up. See the case: readmission prediction

  • 95% fewer data processing errors
    Global CRO · Canada

    Multi-omics data warehouse for drug discovery

    86% of data processing automated. See the case: multi-omics data warehouse

  • 100% IP protection for shipped models
    Biotech · Poland

    Obfuscated, containerized delivery of proprietary models

    Packaging three times faster. See the case: secure model packaging

  • 37 molecules confirmed active in the wet lab
    Series B biotech · USA

    Generative molecule design for skin rejuvenation

    Preclinical narrowing 10× faster than the baseline. See the case: skin rejuvenation drug discovery

Alternatives

Regulatory consultancy, eQMS platform or dev shop? Who fixes the code itself?

A consultancy writes the strategy, a platform stores the evidence, and a dev shop ships features. The gap in between is code that does not match its own documents. That is where we work.

Blackthorn regulated AI, a regulatory consultancy, an eQMS platform and a general software development company compared
CriterionBlackthorn regulated AIRegulatory consultancyeQMS platformGeneral dev shop
Changes the codeYes: refactoring, tests, traceabilityNo, advises onlyNo, stores documentsYes, without regulatory files
AI and model depthPhD scientists and ML engineersRarely in-houseTemplates onlyVaries by team
Writes the evidenceLifecycle, risk and validation filesStrategy and submission textWorkflows you fill inSeldom
Reviewer questionsEngineering answers, with evidenceRegulatory correspondenceNot involvedNot involved
After submissionRetainer through certificationHourly, on requestSubscription continuesA new project

Comparison reflects typical provider scopes, not specific companies.

Why Blackthorn

Engineers who write the code and the evidence for it

  • One team, two outputs. The people who refactor the code write its lifecycle and risk files.
  • Model failure as a hazard. Drift, bias and out-of-distribution input go into the ISO 14971 analysis.
  • Works with your RA. We supply engineering evidence to your regulatory lead or consultancy.
  • No platform lock-in. Files go into the eQMS or document system you already use.
  • Clear lines. FDA and notified bodies decide; we make the product and its file ready for them.

Who is on your regulated project?

A lead engineer owns the rework plan, a scientist validates the model and its datasets, and our QA engineers build traceability and tests with your quality team.

  • Alex Gurbych, PhD, Co-founder and Director
  • Serhiy Protsenko, DrSci, Prof., Principal Data Scientist
  • Dmytro Krasnienkov, PhD, Principal Scientist, Bioinformatics
  • Ivan Izonin, PhD, DrSci, Scientific Advisor, Artificial Intelligence
Meet the scientists

Client reviews

Teams in healthcare and life sciences on working with us
22 reviews, 5.0 on Clutch

  • Verified on Clutch

    «The team surpassed expectations on timelines»

    …provided much needed guidance and overall input on design, all while operating with a high degree of autonomy.

    Carl Kaub, Vice President of Chemistry, HTG Molecular Diagnostics

  • Verified on Clutch

    «Blackthorn.ai is great at adapting to our needs»

    It’s not easy to work with us as a critical infrastructure provider because we have many restrictions that make it difficult to work with data since it has to be secured.

    Under NDA, Deputy, Surgical Center

  • Verified on Clutch

    «What impressed us the most»

    …was their ability to take over a fragile frontend in production, quickly fix a critical authentication issue, and then deliver a large UI redesign with strong QA discipline, all while keeping scope controlled and avoiding risky backend changes.

    Egoitz Arruti, CTO, Ubikare

FAQ

Regulated AI and SaMD questions, answered

01 What is software as a medical device (SaMD)?

Software as a medical device (SaMD) is software intended for a medical purpose, such as diagnosis or treatment, that works without being part of a hardware device. Because it is regulated as a device, it needs a documented lifecycle, risk management and clinical evidence before it reaches patients.

02 What does a medical device software development company do?

It designs, builds and maintains software that is a medical device or runs inside one, and produces the records regulators expect along the way: a development plan, requirements, architecture, verification, risk management and release records under IEC 62304 and ISO 14971. Blackthorn AI works on the part AI products most often lack: making existing code and models match their documentation, then writing the lifecycle, risk and validation files with your quality and regulatory team.

03 Can you take over an AI product built without IEC 62304?

Yes, and it is the most common starting point. IEC 62304 does not demand a rewrite. It demands that the software you ship is planned, specified, verified and under control, and that the records show it; Amendment 1 (2015) added a route for legacy software. Our 2–3 week audit assigns a software safety class, lists third-party components as SOUP (software of unknown provenance) and ranks the gaps a reviewer would flag. Rework then adds requirements, architecture records, tests and traceability around the code you already have, and refactors only the parts that cannot be justified.

04 What does IEC 62304 compliance require from a dev team?

A documented software life cycle scaled to the software safety class, A, B or C: a development plan, software requirements, architecture, verification and testing, a controlled release, and processes for maintenance, risk management, configuration management and problem resolution. The higher the class, the more design detail and verification it needs. Every requirement must trace to its tests and to the risk controls from ISO 14971. Day to day, that means tickets linked to requirements, a versioned SOUP list, code reviews against acceptance criteria and a record of each of these steps.

05 What is a design history file (DHF)?

The record of how a device was designed: plans, inputs, outputs, reviews, verification, validation and design changes. Since 2 February 2026 FDA’s Quality Management System Regulation (QMSR) incorporates ISO 13485, which calls it the design and development file, but reviewers expect the same content. For software it holds the IEC 62304 lifecycle records, the ISO 14971 risk file, the traceability matrix and the validation reports, which are the files we write with you.

06 Will the EU AI Act cover our AI medical device?

Yes, if the AI is part of a device that needs a notified body under MDR or IVDR. Such a device counts as high-risk under Article 6(1) of the AI Act, and the Digital Omnibus on AI (Regulation (EU) 2026/1744) sets those duties to apply from 2 August 2028 for products covered by Annex I, which include medical devices. The AI Act does not change the device’s MDR or IVDR class, and the MDCG 2025-6 guidance recommends one set of technical documentation and one quality management system for both laws. We add the AI Act items, such as data governance, logging, human oversight and accuracy testing, to the files you already keep for MDR.

07 What does FDA expect from AI-enabled device software?

The same premarket evidence as other device software, plus detail on the model. FDA’s draft guidance of January 2025 on AI-enabled device software functions asks for a description of the model and its inputs and outputs, how training and test data were collected, annotated and kept apart, performance testing that includes subgroups, an analysis of bias, a plan for monitoring performance after launch, and labeling that explains the model to its users. Connected devices also need the cybersecurity evidence required by section 524B of the FD&C Act.

08 What is a predetermined change control plan (PCCP)?

A PCCP is part of a marketing submission that describes in advance which changes you plan to make to an AI-enabled device after authorization, how each change will be developed and validated, and how its impact will be assessed. Once FDA authorizes the plan, changes made within it need no new 510(k), De Novo or PMA. FDA first finalized its guidance on PCCPs for AI-enabled device software functions in December 2024. We write the modification protocol from the way your model is actually retrained, so the plan matches the pipeline.

09 What does FDA require for medical device cybersecurity?

For a cyber device, meaning one that includes software, can connect to the internet and could be vulnerable, section 524B of the FD&C Act requires a plan to monitor and fix vulnerabilities after launch, processes that give reasonable assurance the device is secure, and a software bill of materials (SBOM). FDA’s premarket cybersecurity guidance also asks for a threat model, security architecture views, a security risk assessment and test reports. We produce these from the code itself: the SBOM from the build and the threat model from the architecture.

10 How long does it take to make software submission-ready?

The audit takes 2–3 weeks and ends with a rework estimate for your product, which is the only reliable number. The rest depends on the safety class, how much documentation exists and how much the code has to change. Review time comes on top: FDA’s goal for a 510(k) is a decision within 90 FDA review days, while De Novo requests and notified body reviews under MDR usually take longer. The retainer covers that period, including answers to reviewer questions.

11 How much does medical device software compliance work cost?

The 30-minute fit call is free. The code and file audit is a fixed fee agreed before you sign: from €5,000 in Europe, £4,700 in the UK and $6,000 in other regions, excluding VAT where it applies. Rework and documentation are quoted per work package from the audit’s gap list, and the certification retainer is priced monthly once the submission plan is known.

12 How does 21 CFR Part 11 apply to AI in GxP systems?

21 CFR Part 11 sets FDA’s rules for electronic records and signatures that replace paper under GxP regulations. If an AI model creates, changes or signs such records, for example by classifying batch data or drafting a deviation report, the system around it needs validation for its intended use, secure audit trails with time stamps, access controls, record retention and electronic signatures linked to their records. The model itself must be version-controlled, so each output can be traced to the model version and data that produced it. In the EU, GMP Annex 11 sets similar rules, and a new Annex 22 on AI was published in draft in July 2025.

13 What is the difference between CSV, CSA and GAMP 5?

Computer system validation (CSV) is the long-standing practice of proving that a computerized system does what it is meant to do, usually with heavily scripted tests and documents. Computer software assurance (CSA) is FDA’s risk-based approach for production and quality system software, finalized in September 2025: effort follows the risk to patients and product quality, and unscripted testing is acceptable where risk is low. GAMP 5 (second edition, 2022) is the ISPE industry guide for GxP computerized systems; it favors critical thinking and risk-based assurance and has an appendix on AI and machine learning. We follow your validation procedures and apply the risk-based approach to AI components.

14 Do you prepare the 510(k) or MDR submission yourselves?

No. Your regulatory affairs lead or regulatory consultancy owns the strategy, the submission and the contact with FDA or the notified body. We supply the engineering content: software documentation, risk and validation files, cybersecurity records and technical answers to reviewer questions. If you have no regulatory lead yet, appoint one before the submission stage; we work alongside whoever you choose.

15 How do you validate an ML model for a regulator?

Against its intended use, on data it has never seen. We fix the intended use and acceptance criteria first, then test on a held-out dataset kept apart from training, ideally from other sites or devices, and report clinically meaningful metrics such as sensitivity and specificity with confidence intervals. Results are broken down by subgroups such as age, sex, site and device to find bias. The report also states limits of use, known failure modes and how performance will be monitored after release, in line with the Good Machine Learning Practice principles of FDA, Health Canada and the UK MHRA. Failure modes such as drift and out-of-distribution input go into the ISO 14971 risk file.

16 How is this different from your AI Governance service?

Regulated AI covers one product, a medical device, SaMD or GxP system that a regulator will review, and ends with a design history file and reworked code. AI Governance covers every AI system a company uses and ends with a register, policies and controls for internal audit, buyers and the EU AI Act. Device makers often need both; we build each control once and map it to both.

Bring one product. Leave with a gap list.

Share the repository, intended use and whatever files you have. We tell you what a reviewer would reject first and what it takes to fix.