Octave builds multi-tenant AI foundation platform on Amazon Bedrock

Octave provides Reliance, a quality management platform for regulated industries such as medical devices, pharma, and food manufacturing, serving over 650 enterprise customers across 180 countries. Data isolation requirements are strict: complaint records, audit findings, and corrective actions cannot leak across tenants. Rather than building a single AI feature first, the team built a horizontal foundation platform, then vertical features like complaint summarization and triage; the platform cut deployment time per use case by 50%. The three-layer architecture: a data preparation layer writes documents via Cohere Embed v4 on Amazon Bedrock into tenant-isolated Amazon OpenSearch Service vector stores; a secure execution layer authenticates through AgentCore Gateway, then routes requests to isolated microVM sessions calling Claude models; a content protection layer applies Bedrock Guardrails on both input and output. Each user session runs in a dedicated microVM that stops and clears memory when the session ends, and credentials are scoped per user and agent and valid only during the session. Embed v4 places images and text in the same vector space and natively supports multiple languages, eliminating two preprocessing pipelines. Based on production query testing, the team reduced output dimensions from 1536 to 512, meeting accuracy targets while cutting storage ~66%. For platform teams that must import massive tenant data within one business day and keep data in-region, these tradeoffs directly define cost and compliance boundaries.