InsFocus AI Platform
Multi-tenant hosted AI service for insurance BI customers. Sits on top of the on-prem InsFocusBI platform.
Architecture
- Portal — customer-facing frontend.
- Auth — OAuth2 client credentials.
- Data — one database per tenant.
- Retrieval — SQL Server 2025
VECTORtype with hybrid search (RRF). - Billing — credit-based.
- Design doc:
insfocus-ai-services-design.md.
Lifecycle
- CI/CD in GitHub Actions. Push to
developHelm-deploys toai-devon Dev MicroK8s. See AI platform deployment. - Release pipeline runs Semgrep OSS + Trivy. See GitHub Actions.
Configuration
- Dev URLs:
ai-portal-dev.insfocus.com(portal),ai-api-dev.insfocus.com(api gateway). - Portal auth: Cloudflare Access (Azure AD) in front; the portal validates the Access JWT and maps the email to a
PlatformUser. - Api auth at the edge: Cloudflare Access service tokens only.
- Portal BFF serves the Angular SPA from
ai/portal/wwwroot/dist, with SPA fallback;/api/**misses return 404.
Decisions
- Backfill — Per-tenant database isolation.
- Backfill — Credit-based billing.
- Backfill — OAuth2 client credentials for API auth.
- Backfill — Vector search on SQL Server 2025
VECTOR, hybrid RRF retrieval. - Backfill — UI designs made in Claude Design.
- 2026-08-22 — Hostnames
ai-portal[-dev].insfocus.comandai-api[-dev].insfocus.com. Why: single-label names are covered by Universal SSL. - 2026-08-22 — Portal trusts Cloudflare Access (Azure AD) as its login in deployed envs. Why: avoids a second OTP login.
Open
- LLM orchestration: Microsoft.Extensions.AI vs Semantic Kernel Process Framework — being evaluated.
- Local LLM for data-residency-sensitive steps — being evaluated.
- Azure AI Foundry for customer-facing deployments — being evaluated.