Octave builds multi-tenant AI foundation on Amazon Bedrock for 14,000 customers

Octave, maker of the Reliance quality management system, built a shared AI foundation on Amazon Bedrock before shipping any single AI feature. According to AWS, the approach has cut deployment time per use case by 50 percent while keeping each customer's regulated data isolated.
- Octave built a shared AI foundation on Amazon Bedrock before shipping any single feature, according to AWS.
- The foundation has three layers: data preparation, secure execution, and content protection with Bedrock Guardrails.
- Cohere Embed v4 at 512 dimensions cut storage roughly 66 percent versus 1,536 while holding target accuracy.
- Reliance hosts more than 2 billion documents, and the team aimed to onboard its largest tenant within one working day.
- AWS says the foundation cut deployment time per use case by 50 percent.
Why Octave built a foundation first
Quality management in regulated industries depends on history: a similar defect investigated last year, a Corrective and Preventive Action (CAPA) that fixed a root cause, or an audit finding that flagged the same gap. According to the AWS post, that knowledge sits in billions of documents that no human can search by semantic similarity across languages and formats, so every investigation starts from scratch.
Octave operates across 180 countries and serves more than 60% of the Global Fortune 500. Its cloud-native system, Reliance, supports over 650 enterprise customers with over 40 applications, and the broader platform serves over 14,000 customers across medical devices, pharmaceuticals, and food manufacturing. Each customer is a separate tenant, and no data can cross a tenant boundary. A single inadvertent disclosure, the post says, would carry regulatory consequences far outweighing any AI benefit.
Workshops with AWS surfaced four use cases: complaint summarization and triage, audit readiness monitoring, incident narrative generation, and predictive quality risk scoring. Because the platform has dozens of applications with different document types and workflows, building AI for one in a silo would mean rebuilding infrastructure for the next. The team therefore built the horizontal platform first and cut vertical feature slices into it. Complaint summarization and triage went first to prove the architecture.
Four tensions and a three-layer design
Before choosing a model, the team worked through four problems that pulled in different directions:
- Data availability without disclosure: vector store, retrieval logic, and generation context all had to be scoped to a single tenant.
- Isolated agent execution: agent permissions had to be scoped to the requesting user rather than a shared service role. A shared agent pool with static credentials failed this test.
- Untrusted content: agents ingest complaints, email attachments, scanned PDFs, and images from outside the organization. The post cites OWASP Top 10 for LLM Applications (2025) risks including prompt injection (LLM01), sensitive information disclosure (LLM02), improper output handling (LLM05), and excessive agency (LLM06).
- Scale: Reliance hosts more than 2 billion documents, and the team committed to onboarding even the largest tenant within a single working day without trading away quality.
The resulting foundation has three layers. Data preparation embeds documents with Cohere Embed v4 on Amazon Bedrock and writes them to a per-tenant vector store on Amazon OpenSearch Service. Secure execution passes requests through AgentCore Gateway, which applies authentication and policy before handing them to an isolated microVM session that calls an Anthropic Claude model on Amazon Bedrock. Content protection wraps both, applying Amazon Bedrock Guardrails for input filtering and again for output filtering before anything reaches the handler. Retrieval, execution, and screening stay in separate stages so each can be enforced and evidenced on its own.
Why Cohere Embed v4
The team tested embedding models against its own corpus rather than public leaderboards. Quality records mix photos of defective products, scanned inspection forms, and annotated diagrams with text. According to the post, Embed v4 places both modalities in one vector space, so a photo of a cracked housing and the paragraph describing it can be retrieved from a single query. Native multilingual support removes the need for translation preprocessing, which matters because documents often arrive in several languages within one tenant.
Region residency was a precondition, since customer data cannot leave its AWS Region, even for inference. Embed v4 supports in-Region inference in US East (N. Virginia), Europe (Ireland), and Asia Pacific (Tokyo), covering three of Octave's regional compliance requirements. The post notes that availability varies by Region and changes over time.
On cost, OpenSearch Service expenses scale linearly with dimension count. Embed v4 offers configurable output dimensions of 256 to 1,536. In tests on representative production queries, 512 dimensions held target accuracy while cutting storage roughly 66 percent against the 1,536 default. The 128K token context window let the team pack many chunks of roughly 4K tokens into each request, which the post credits for the onboarding throughput.
Isolation in the runtime, and what it means
Octave chose to enforce tenant separation in the runtime rather than in application code, which would have required every future feature team to get it right on every code path. AgentCore runtime enforces separation below the application layer. Because it is framework- and model-independent, the choice of agent framework and models for the Complaint Triage Assistant was not constrained by the isolation decision. The source text breaks off while describing how each user session runs in a dedicated environment.
For other SaaS providers, the post argues the reasoning applies beyond quality management wherever data is valuable but isolation is non-negotiable. The reported outcome of the foundation-first strategy is a 50 percent cut in deployment time per use case, a figure given by AWS in its account of Octave's work.


