Trust Center | Rendex — Private AI for Law Firms

Trust Center

Forwardable proof artifacts and architecture details for IT, security review, and firm leadership.

View Proof Pack Book a Demo
Architecture Proof Pack Deployment Model Governance Trust Controls Evaluation Checklist Technical Appendix Capacity FAQ
Architecture

Architecture at a glance

What runs inside the confidential enclave, what's exposed, and how retrieval works — without implementation-level detail.

Dedicated Azure confidential VM

The full Rendex stack runs in containers inside a dedicated Azure confidential VM provisioned for your firm. Sealed by AMD SEV-SNP memory encryption and backed by an NVIDIA H100 TEE.

Only 80/443 exposed

TLS reverse proxy handles all inbound traffic. All other services are inside the confidential VM's private network and not reachable from your LAN.

Hybrid retrieval

Vector search + full-text search + reranking, with ACL filtering applied before context assembly. Users only see documents they have access to.

Verifiability

Page-level citations on every answer. Exportable audit logs for every query. Both are available for your team to inspect during the evaluation.

Detailed versions, internal ports, and scripts are in the Technical Appendix.
Hardware platform
InstanceAzure NCCadsH100v5 (Confidential Computing)
CPUAMD EPYC with SEV-SNP memory encryption
GPUNVIDIA H100 80 GB HBM3 with TEE attestation
MemoryHardware-encrypted (SEV-SNP sealed)
StorageEncrypted managed disks
OSUbuntu 24.04 LTS confidential VM image
RegionCustomer-selected (US, EU, or CA)
AttestationAMD SEV-SNP quote + NVIDIA H100 TEE report on demand
Core services
  • Application server / API / web UI
  • Reverse proxy with TLS termination
  • LLM inference engine
  • Vector database (semantic search)
  • Full-text search engine (keyword / BM25)
  • Relational database (auth, audit, metadata)
  • Session cache
  • Document parsing, OCR, and vision-language extraction
RAG pipeline (high-level)
Ingest Parse / Vision OCR Chunk Embed Index Retrieve Rerank Generate Cite

ACL filtering is applied at the Retrieve step, before any document context is assembled or sent to the language model. Users cannot retrieve documents outside their authorized matters.

Networking model

Inbound: Ports 80 (redirect) and 443 (HTTPS) only. All other services communicate on an internal-only container network.

Outbound: No outbound egress required after initial setup. Verifiable via the Enclave Attestation Kit and Azure NSG flow logs.

Internal port assignments are documented in the Technical Appendix.

Proof Pack

Verification Artifacts

Documents your team can forward to IT, security, and firm leadership. Evaluation deliverables are generated from your environment.

Permission Parity Report

User A vs User B access tests + canary phrase checks. Demonstrates fail-closed enforcement.

Template (example format) ↓

Citation Accuracy Report

Verified page-level citations across 30–100 queries. Tracks incorrect/ambiguous cites and refusals.

Template (example format) ↓

Enclave Attestation Kit

AMD SEV-SNP attestation quote + NVIDIA H100 TEE report proving your enclave is hardware-sealed.

Template (example format) ↓

Audit Export Sample

CSV/JSON export examples: query → sources/pages → response metadata, filterable by date/matter/user.

Template (example format) ↓
Request Proof Pack Book a Demo
Templates show the format and structure of each artifact. Actual proof packs are generated from your environment during an evaluation.
Operations

Deployment & Operations

Same-day provisioning with minimal disruption and clear operational ownership.

Same-day enclave provisioning + smoke test

Pre-configured confidential VM. We provision the enclave, your IT runs the smoke test, and confirms all services are healthy. No hardware to ship. Typical setup: same day.

Hardware-enforced isolation — verified at boot

A fresh AMD SEV-SNP attestation quote and NVIDIA H100 TEE report are generated on every boot. All container images and model weights are sealed inside the enclave. No internet required for core workflows after initial setup.

Auto-recovery after restart

All services configured to start automatically when the VM reboots. Attestation is re-verified on every boot. No manual intervention required after a restart.

Health monitoring

Built-in health endpoints for GPU utilization, disk usage, memory, and service latency. Your monitoring stack can poll standard HTTP health checks.

Rolling updates with rollback

Updates are versioned container images applied via secure channel. Rollback to previous version documented and tested. No forced migrations.

Model Governance

Model governance

Rendex is model-agnostic. Deployments use only customer-approved models, with a license and provenance summary for each.

TierPolicyExamples
ApprovedCommercial-permitted + customer-approvedGPT-OSS (default)
OptionalSupported on request; subject to policy + license reviewAlternative open-weight models
Not allowedNon-commercial or restricted licenses for paid deploymentsAny “research-only” weights
Model license summaries are available on request. No model weights are transmitted outside the enclave.
Security & Compliance

Trust controls

Summary of controls relevant to security review.

TLS via reverse proxy; same-origin policy

All client traffic encrypted via HTTPS (TLS 1.3). Self-signed certificate replaceable with your CA. Same-origin policy enforced on the web UI.

ACL-enforced matter isolation at retrieval time

Document access is filtered at the retrieval layer before context assembly. Users cannot retrieve documents outside their authorized matters, even with direct queries.

Append-only audit log with export

Every query, login, and document action logged to an append-only table. Exportable as CSV for compliance review. The database rejects UPDATE and DELETE operations on audit records.

Zero third-party data exposure — verified at four layers

Memory encrypted by AMD SEV-SNP. GPU sealed by the NVIDIA H100 TEE. No outbound egress after setup. No telemetry, no license checks, no phone-home. All four layers verifiable via Azure Attestation Service reports.

SSO via OIDC (Microsoft Entra ID)

OpenID Connect with JWKS signature validation. Your firm's MFA, Conditional Access, and device compliance policies apply automatically.

Rate limiting on API endpoints

Configurable rate limits on all API routes. Protects against abuse and ensures fair resource allocation across users.

Vendor questionnaire turnaround: 48 hours

Send us your security questionnaire and we return it completed within two business days. No NDA required for the initial packet.

SOC 2 aligned architecture

Built around SOC 2 control objectives: access control, audit logging, encryption, change management. Not yet independently audited — we provide verification steps so your team can assess directly.

Evaluation

30-Day Evaluation Checklist

Concrete metrics your team can measure during the evaluation — not vendor promises.

  1. Time saved by workflow. Measure hours saved on chronology, clause table, and precedent search tasks against your current manual baseline.
  2. Citation coverage rate. What percentage of generated answers include at least one source citation? Target: 100% of factual claims.
  3. Citation correctness (spot checks). Pull 20–30 cited answers at random. Verify the cited passage actually supports the claim.
  4. Matter isolation tests. Confirm users assigned to Matter A cannot retrieve documents from Matter B via cross-matter queries.
  5. Audit export review. Export the full query log. Confirm it captures user, timestamp, query text, retrieved sources, and generated response.
  6. Enclave attestation review. Request the AMD SEV-SNP attestation quote and NVIDIA H100 TEE report. Confirm your enclave is running the expected code inside a hardware-sealed confidential VM.
Appendix

Technical Appendix

Technical details are provided for evaluation. Internal service ports are not exposed externally; only 80/443 are reachable from your network.
Container inventory
ServiceRoleInternal only
NginxTLS reverse proxy, rate limitingPorts 80/443 exposed
Chat UI (Node.js)Web interface + API serverYes
GPU Inference EngineLLM + embedding inferenceYes
QdrantVector databaseYes
OpenSearchFull-text / BM25 searchYes
PostgreSQLAuth, audit, metadataYes
ValkeySession cacheYes
TikaDocument parsing + OCRYes
Vision LMMachine vision OCR for scanned PDFsYes
Internal ports (not externally exposed)
ServiceInternal portProtocol
Chat UI3000HTTP
GPU Inference11434HTTP
Qdrant6333 / 6334HTTP / gRPC
OpenSearch9200HTTP
PostgreSQL5432TCP
Valkey6379TCP
Tika9998HTTP

All ports above are bound to the internal Docker bridge network. Only 80/443 are reachable from the host network.

RAG pipeline parameters
ParameterValue
Chunking strategyRecursive, ~512 tokens with overlap
Embedding modelConfigurable (default: bundled inside the enclave)
Vector top-k20 (configurable)
BM25 top-k20 (configurable)
RerankerCross-encoder, top-5 after rerank
ACL filteringApplied at retrieval, before context assembly
Citation methodPage-level, linked to source document viewer
Database schema (overview)
  • users — SSO identity, role, created/updated timestamps
  • matters — matter metadata, access control lists
  • documents — file metadata, matter association, ingestion status
  • chunks — text segments with document/page references
  • audit_log — append-only: user, action, query, sources, response, timestamp
  • sessions — chat history, user association
DMS integration status
SystemStatus
iManageAvailable now
NetDocumentsIn development
SharePointIn development
Drag-and-drop uploadAvailable now
Capacity & Performance

How Capacity Works

A single enclave is designed for 5–15 concurrent active users, and can support up to ~20 with tuning. Private Cloud Standard (shared H100) covers most small to mid-size firms. Private Cloud Dedicated (single-tenant H100) provides priority inference for higher-volume firms up to around 40 attorneys.
During usage spikes — when many people ask questions at once — Rendex briefly queues requests so the system stays stable. Answers may take a bit longer to appear, but the system remains responsive and fast enough for normal legal workflows.
We recommend Private Cloud Dedicated when you have 50+ attorneys using Rendex regularly, when you need the full 120B model, or when you expect high concurrent usage and want priority inference with no queue.
Yes. For the best experience, we suggest scheduling large bulk ingestion for evenings or weekends, or throttling big uploads during peak hours, so day-to-day queries stay responsive while documents are being indexed in the background.
FAQ

Common Questions

Rendex runs inside a dedicated Azure confidential VM, sealed by AMD SEV-SNP memory encryption and the NVIDIA H100 trusted execution environment. Not even Microsoft can read your data inside the enclave. The Trust Center shows how to verify controls and outputs independently via cryptographic attestation reports.
Request the AMD SEV-SNP attestation quote and the NVIDIA H100 TEE report. These cryptographic proofs confirm your data and model weights are sealed inside a hardware-attested confidential VM that no operator — including Microsoft — can read.
Yes. Email info@rendex.inc and we'll return it completed within 48 hours. No NDA required for the initial packet.
Not yet. The architecture is aligned with SOC 2 control objectives (access control, audit logging, encryption, change management), but we have not completed independent third-party audit. We provide verification steps so your team can assess directly.
Rendex is model-agnostic. Default deployments ship with GPT-OSS (OpenAI 120B MoE, commercially licensed). Alternative models are supported on request after license review. No model weights are transmitted outside the enclave.
Platform

Platform Capabilities

Full audit trail (queries, logins, exports)
DMS sync (iManage workspaces)
Bulk ingest at scale (100K+ docs)
Matter summarization (one-click)
Document comparison + memo export
Guided workflows (depo prep, chronology, due diligence)

Ready to review?

We respond within 2 business days. No NDA required for the initial proof pack.

Request Proof Pack Book a Demo