You're seeing this page as if you were . The main menu is still yours, though. Exit from immersion
Mathilde De RoumillyMD

Mathilde De Roumilly

LLM RAM Forensics · SHA-256 Amnesia Proof

€1,100/day
Auray, FR
0-2 years

Average response time: 1 hour

Freelancer profile translated to English.
Back to original language

About Mathilde

Your LLM processed sensitive data. It claims to have purged. Who proves it?
I am Mathilde De Roumilly, founder of Scanalis, the first LLM RAM forensic audit service in France.
The Canary Protocol injects 8 cryptographic tokens into your AI system via its API standard, triggers the purge, scans RAM byte by byte across 4 zones. Binary verdict: AMNESIA_CONFIRMED or AMNESIA_FAILED. SHA-256 sealed PDF report, RSA-PSS signed, RFC3161 timestamped. Enforceable by CNIL and sector regulators.
Non-intrusive by design. Zero agents installed. Zero SSH access. Zero source code access. You will generate the memory dump yourselves. I scan. I sign.
I work with CIOs, DPOs, and CISOs who need physical proof, not a contractual declaration that their LLMs truly forget personal data after processing.
3 slots per month. I scan. I sign.

Service 1: Canary Diagnostic
Title: LLM RAM Forensic Audit: Canary Protocol · SHA-256 Report
Description: Did your LLM physically forget? I prove it.

Service 2: Canary Audit
Title: Full RAM Forensic Audit: Canary + GPU + RFC3161 + Chain of Custody
Description: The complete proof. Enforceable against regulators, insurers, investors.

Service 3: Minerva Import
Title: LLM Memorization Audit: Minerva Protocol · Sensitive Zone Import Decision
Description: Before importing a model into a sensitive, air-gapped, or classified zone, know what it carries.

Service 4: Complete AMNESIA Project
Title: Full Support: From Blind Spot to AMNESIA_CONFIRMED Certification
Description: We don't let go until it's AMNESIA_CONFIRMED.
  • French

    Fluent

Remote only
Primarily works remotely

Experience

  • Anonyme
    LLM RAM Forensic Audit - Post-Purge Residue Detection · Digital Health Sector
    BIOTECH
    March 2026 - March 2026
    Rennes, France
    Publisher of a medical AI solution deploying an on-premise LLM for patient data processing.

    Execution of the Canary Protocol: injection of 8 cryptographic tokens targeting health data categories (Art.9 GDPR), patient identifiers, contact information. Triggering of the declared system purge procedure. Forensic scan of the 4 memory zones. Duration: 5 days.

    AMNESIA_FAILED: 3/8 tokens detected in RAM.
    Residues located in the Python application heap and Linux page cache. Exact memory addresses documented in the report. Exposed category: health data (tokens 2, 5, 7). Residues persisted 22 minutes after session closure and declared purge.
    Identified cause: the purge procedure called del() on Python session objects without physically overwriting bytes. The Linux page cache was not released via posix_fadvise().

    Habitable Phase 1
    13-page forensic report, AMNESIA_FAILED verdict, hex dumps with memory addresses, IPE (Index of Probability of Exposure): 74/100.
    SHA-256 sealed, nominally signed. Transmitted to the CIO and DPO within 72 hours.

    Phase 2 Deliverable: Remediation
    Technical correction specifications: implementation of SecureBuffer with ctypes.memset() 3 passes, page cache release via posix_fadvise(POSIX_FADV_DONTNEED), mlockall() configuration to disallow swap. Technical sheets provided for the client's exact stack.

    Return to Phase 4 after implementation of corrections for AMNESIA_CONFIRMED certification delivery.

    The client avoided an Art.9 GDPR health data exposure during the planned regulatory audit. Without this audit, the purge would have been presented as effective based solely on the DPA, without any physical proof.

    Python · RAM Forensic Analysis · Canary Protocol · ctypes · posix_fadvise · SHA-256 · Heap Analysis · Linux Page Cache
    RAM Memory Forensics SHA-256 RSA-PSS Cryptography
  • Anonyme
    LLM RAM Forensic Audit - Offline Autonomous System · Tech Sector
    CONSULTING AND AUDITS
    February 2026 - February 2026
    Auray, France
    Client from the technology sector deploying an LLM on sensitive data in an offline environment. Autonomous system without external network connection. The client had a declared purge policy but no means to verify its physical execution.

    Execution of the Canary Protocol in offline mode: injection of 8 cryptographic tokens via the system's local API, triggering the purge procedure, forensic scan of the 4 memory zones (application heap, page cache, swap, kernel buffers). Duration: 5 days.

    AMNESIA_CONFIRMED: 8/8 tokens purged.
    The system physically destroys data in all 4 zones. 13-page PDF report delivered, SHA-256 sealed, nominally signed. The client now has legally enforceable forensic proof for the CNIL and sector regulators that their purge procedure is physically effective.
    Technologies
    Python · RAM Forensics · Canary Protocol · SHA-256 · RSA-PSS · RAM Analysis
    RAM Memory Forensics RAM Analysis

Recommendations

Be the first to recommend Mathilde

Help this freelancer shine by sharing your experience working together.

These freelancer profiles also match your criteria

AgathaA

Agatha Frydrych

Backend Java Software Engineer

4.7

(3)

2

BaptisteB

Baptiste Duhen

Fullstack developer

4.6

(4)

5

AmedA

Amed Hamou

Senior Lead Developer

4

(2)

7

AudreyA

Audrey Champion

Web developer

4.3

(3)

4

Categories