Cloud Native Thane - July Meetup

AI workloads introduce a new class of security challenges that most cluster operators aren't ready for: long-lived API keys for model providers, multi-tenant model servers with no request isolation, and ML artifacts with no supply chain verification.

This talk applies zero-trust principles specifically to AI infrastructure: using SPIFFE/SPIRE for cryptographic identity between services and model servers, short-lived credentials via cert-manager to replace hardcoded LLM API keys, OPA/Gatekeeper policies for model endpoint access control, and Cosign for signing and verifying ML model artifacts.

Each pattern is demonstrated with a working Kubernetes setup, and common misconfigurations are called out explicitly.

Modern organizations process thousands of unstructured documents every day—from financial statements and reports to invoices and disclosures. Building a reliable system to ingest, process, and transform this data is less about machine learning and more about designing resilient cloud-native platforms.

In this talk, we'll explore a practical architecture for processing documents at scale using cloud-native patterns. We'll walk through a layered pipeline that separates ingestion, orchestration, processing, indexing, and data modeling into independent services, making the platform easier to scale, operate, and evolve over time.

Rather than focusing on a specific business domain, we'll discuss architectural decisions, trade-offs, and operational lessons learned while designing distributed document-processing systems.