Compex
Request a Pilot

Hardware-Enforced Runtime for Cloud VPC & On-Premise

Run LLMs, Agents, ML Models, and Clean Rooms on Data You Can’t Move.

Compex executes high-trust workloads inside hardware-attested Trusted Execution Environments (TEEs) right next to sensitive records. Lock any self-hosted LLM, autonomous agent, custom ML pipeline, or multi-party clean room to a signed Pipeline Contract at the silicon level—and auto-generate the cryptographic evidence bundle DPOs and regulators require.

See the sandboxBook a technical walkthrough
A one-minute film. AI is ready for real data, but pipelines can run anywhere and nobody can say who signed off. Compex is a hardware-enforced runtime for regulated AI: approve the scope before any LLM, agent or ML pipeline runs, execute it inside an attested Trusted Execution Environment while out-of-scope operations are refused, then verify a hardware-attested hash chain of every event. Built for insurance, banks, healthcare and research.

Data access is where deployment stops.

AI teams need real data. Data owners have no way to check what happens to it.

Data

Sensitive records cannot be copied into every model, tool or vendor environment.

Legal

Review teams need clear policy, approval and audit evidence before work can start.

Proof

AI teams need a way to run inference that survives a vendor security review.

The same governed path, whatever the field names are

  • GDPR Art. 9
  • Special-category fields

Insurance and Insuretech

Diagnoses, addresses, payout history. Every field is classified in the contract before a workload sees any of it.

01/04
Open the Insurance and Insuretech sector brief

Compex creates a controlled path between data owners and AI teams.

Ownership, execution and approval stay separated.

Protected data

records · scans · claims

stays put

Cryptographic Pipeline Contract

Data owners and DPOs sign off on exact column and row scope, command allow-lists and output limits. The contract hash is bound directly to the enclave before boot.

TEE-Isolated Runtime

The workload and policy gate run inside an encrypted hardware enclave (CPU/GPU TEE). The host OS, cloud admin and AI vendor cannot inspect plaintext memory or alter policy rules mid-run. Out-of-scope calls are killed at runtime.

Hardware-Attested Evidence

The enclave outputs an ordered hash chain of every allowed and blocked operation, paired with a hardware remote-attestation quote proving the exact approved code and contract ran without tampering.

job specapproved output

AI team

risk model · scoring API

reviews

DPO / Legal

Reviews scope before the run. Receives signed evidence after it.

Built for European regulation

Where the runtime carries the burden, and where it stays with you.

GDPR

Binds purpose, field and output policy to the run, and records who approved it.

Your duty

Lawful basis, role allocation, transparency, data-subject rights and DPIA conclusions.

DORA

Logs runtime access and deployment, and timestamps every incident.

Your duty

Accountability, regulatory reporting and resilience testing.

EU AI Act

Gates execution behind approval and logs what ran, for high-risk system duties.

Your duty

System classification, conformity assessment, registration and provider/deployer duties.

NIS2

Hashes each execution event and lists the workflows affected.

Your duty

Scope determination, risk governance and statutory notification.

An audit-ready runtime by design.You keep control of your obligations.

One contract. One run. One file.

  1. 01 Pipeline Contract

    Both sides agree the scope in writing, and the reviewer signs it. Nothing executes until that version exists.

    Pipeline Contractv1 approved
    workload
    compex-analyze-duckdb
    field scope
    3 of 6 admitted
    output
    aggregates only
    purpose
    fraud model validation
    dataset
    synthetic-insurance-claims
    retention
    30 days
    reviewer
    named, signed

  2. 02 Governed runtime

    Ask for a field outside the contract and the job is refused mid-run, and the refusal logged.

    1. Pipeline Contract
      approved · v1
    2. Workload admitted
      compex-analyze-duckdb:2026.05.29-r7
    3. Executing in scope
      synthetic-insurance-claims · 100 rows
    4. Out of scope
      if raised: job stops, refusal enters the chain
    5. Evidence sealed
      sha256 526183fa…f59b21

    Refusal is what the runtime does with an out-of-scope operation. The captured sandbox run completed without one.

  3. 03 Evidence bundle

    Everything the run touched, in one file, checkable against the contract that authorised it.

    evidence.json142 KB

    sha256 526183fa…f59b21

    summary.pdf88 KB

    sha256 2c90f220…b1586d

    audit.csv11 KB

    sha256 8a6ac4e1…a000eab

    Evidence bundlesealed

What a reviewer can check without asking us

Generated at execution time and hashed as it goes.

  1. 01Datasetconfigured
  2. 02Policyconfigured
  3. 03Approvalapproved
  4. 04Executioncompleted
  5. 05Evidenceready
Five stages, each naming its owner. The four that change state are hashed into the chain; the bundle is what they produce.

Who signed what

Every stage names who signed it.

first hash

3530a8a90042108515722e83778f8af0d07f3c96f41489817fe66c299c56eabb

last hash

c35620aeaebae1fdb2479cff13413b4e941f3a0496a5c6384ea4deae067d1ce2

events
4
valid
4
problems
0

The chain, as recorded

First hash, last hash, event count. An empty problems array.

evidence.jsonmatch
stored

526183fab2767c8360df9670d0e57334138c0f0849c6f4da77f7d8bf30f59b21

computed

526183fab2767c8360df9670d0e57334138c0f0849c6f4da77f7d8bf30f59b21

summary.pdfmatch
stored

2c90f220d5ca7dcc7ea63fe54a597c4ba67811444fef119375a04a33c4b1586d

computed

2c90f220d5ca7dcc7ea63fe54a597c4ba67811444fef119375a04a33c4b1586d

audit.csvmatch
stored

8a6ac4e1fcb3b9947045e2d8cbbf37cc4c6c287cc9f4605f86254d3c0a000eab

computed

8a6ac4e1fcb3b9947045e2d8cbbf37cc4c6c287cc9f4605f86254d3c0a000eab

Stored against computed

Both hashes in full. Compare them yourself.

Look at a real evidence bundle before you talk to us

Train, Analyze and Serve run under the same governed path. Below is an Analyze job end to end: EU-hosted, synthetic insurance data, isolated tenant. A reviewer approves the scope, the job runs, and the bundle verifies against its own hashes.

Open the sandbox
app.getcompex.com/try

The same run answers to everyone.

AI providers

What do I hand the client's lawyer?

Your model runs against the real data under an approved contract inside a TEE, so the vendor never touches raw customer rows. B2B AI startups use Compex to unblock stalled six-figure enterprise pilots in banking, insurance and healthcare, and hand the client's legal team a hardware-attested evidence bundle.

Data owners

Can I prove the rows never left?

The records never move. You approve the scope, name the reviewer, and keep a record of every operation inside the run. Run self-hosted open-weight models (Llama, GLM, Qwen) and internal coding or claims agents inside a command-allow-listed TEE sandbox with zero external data egress.

DPOs and legal reviewers

Who am I taking at their word?

Nobody—not even the AI vendor or the cloud admin. You verify the signed Pipeline Contract, the hardware TEE attestation and the immutable hash chain of every operation yourself in JSON, PDF or CSV.

Four stages from blocked deal to signed evidence

One use case, one dataset, four stages. You end up with results and the evidence legal needs.

  1. 01AI provider and data owner

    Scope

    One use case, one dataset, the fields you need. That becomes the draft contract.

  2. 02Data owner and DPO

    Approve

    The reviewer reads the scope and approves a version.

  3. 03AI provider

    Run

    Your workload runs.

  4. 04Both parties and legal

    Review

    The bundle is exported and the hashes checked.

Questions we get asked

No. It runs next to the data under the contract's field and output limits. The provider gets the permitted outputs and the bundle. The raw records never move.

The Pipeline Contract declares permitted operations, field scope and output restrictions before execution. The Compex runtime sits inside a Trusted Execution Environment, where memory is encrypted at runtime and isolated from the host operating system and hypervisor. The workload cannot bypass the contract even if the underlying host machine or container orchestrator is compromised. An out-of-scope operation is refused while the job is running, and the refusal is written into the evidence bundle as a blocked action.

Each event and each export file is hashed independently, and the hashes are chained in order. The bundle also carries a cryptographic remote-attestation quote anchored to the underlying TEE hardware (AMD SEV-SNP, Intel TDX or NVIDIA Confidential Computing). That proves the hashes were generated inside a genuine, un-tampered enclave running the exact signed Pipeline Contract. Checksum verification compares stored values against computed ones, and the bundle exports as JSON, PDF and CSV for your reviewer to check directly.

Deploy Compex in your existing cloud VPC on confidential instances (AWS, GCP and Azure Confidential VMs, including Confidential H100 and Blackwell GPUs) or on-premises on air-gapped hardware. Data never leaves your residency boundary.

No. That is your assessment to make. We give you what it needs: the signed contract, the runtime, and the run record.

Duration and retention are fields in the Pipeline Contract, agreed before execution. Output restrictions limit what may leave, and the bundle records what was produced and what was blocked.

The deal is stuck on evidence.

Bring us the one that has been with legal for months. We will show you the file their reviewer gets.

Request a pilotSee the sandbox