AI Architecture & Engineering
Deep dives into multi-tenancy, RAG systems, vector databases, and SaaS engineering.
Guardrails as Tenant Configuration: Beyond One-Size-Fits-All Security
Why globally hardcoded guardrails fail in multi-tenant AI products, how to build input and output checks as a configurable pipeline per tenant, and why guardrail decisions need an audit trail.
Multi-Tenancy for AI & RAG Systems: The Landscape Behind the Hype
Why classical SaaS multi-tenancy breaks at four new frontiers in RAG products, and why caching—semantically and technically—is the most easily-overlooked place for cross-tenant leaks.
Token Costs per Tenant: Why Rate Limiting Works Differently for LLM-SaaS
Why 'requests per minute' is the wrong metric for LLM products, how multi-provider pricing complicates cost tracking, and how the reservation pattern prevents tenants from busting their budget undetected.
Vector-Store Isolation in Multi-Tenant Systems: Where RAG Architectures Really Break
Why the classic tenant_id column doesn't work for vector stores, which isolation models exist (and what they mean concretely in Pinecone, Weaviate, Qdrant, or pgvector), and how to structurally prevent cross-tenant leaks in RAG retrieval instead of hoping no one forgets a filter.