Edge — Production-grade LLM evaluation platform

→ Business impact: From an internal prototype to a platform usable inside a regulated banking environment — SSO propagated, tamper-evident audit, zero critical CVE past the gate.

Context

Inside a European private bank (client under NDA), the team needed a platform to evaluate language models under production conditions — compare, measure, trace — in an environment where auth, audit and reproducibility are non-negotiable.

This page is kept intentionally high-level. Technical details and demo available on request, under NDA.

Technical challenge

SSO propagation from an auth domain to an app domain without re-authentication, respecting the bank’s OIDC constraints.
Ship prod-grade infra: GitOps, reproducible deploys, rollback.
A build chain without critical CVEs and a tamper-evident audit log.

Solution

Auth delegated via oauth2-proxy / OIDC, cross-domain session propagation controlled end-to-end.
Kubernetes deploys driven by GitOps (ArgoCD + Helm) — every release reproducible and reversible.
FastAPI backend + PostgreSQL, audit log with verifiable integrity.
Hardened CI/CD: dependency and image scans, 0 critical CVE past the gate.

Outcome

Cross-domain SSO working in production between the two app domains.
Prod-grade CI/CD: reproducible builds, GitOps promotion, rollback in a commit.
Secure chain: zero critical CVE past the security gate.
Tamper-evident audit live for evaluation traceability.

Learnings & trade-offs

In a banking environment, speed comes from rigour, not despite it. GitOps + automated gates = more confidence to ship often.
Cross-domain auth is the most under-appreciated and most critical part of an internal platform.