Kofi E. Ampaabeng
All work
Designed & builtInternal platform

ai-llm — Secure Multi-Model AI Gateway

A self-hosted gateway that routes a team's AI through one endpoint — local GPU inference with Anthropic/OpenAI fallback — behind PII redaction, injection defense, audit logging, and cost control.

  • LiteLLM
  • vLLM
  • Ollama
  • Presidio
  • Docker
  • Prometheus / Grafana

What it is

A self-hosted corporate LLM gateway. Every engineer's AI request — from the editor, an app, or a chat UI — goes through one OpenAI-compatible endpoint that runs local GPU inference first and falls back to the cloud, with security, governance, and cost control applied on the way through. It's the "secure gateway/proxy that integrates multiple frontier models and is optimized for logging and control" pattern — running, not theoretical.

One endpoint, many models

  • Routes across vLLM (local GPU), Anthropic, OpenAI, and Ollama behind a LiteLLM gateway — local-first to keep cost and data in-house, cloud fallback when needed.
  • Onboard a new frontier model by configuration, not code.

A security & control proxy

  • An ordered guardrail pipeline on every request: PII redaction (Presidio) → prompt-injection defense → engineering-discipline policy → audit logging.
  • API-key provisioning, rate limiting, and per-request cost tracking — so usage is attributable, governed, and bounded.

Observability & operations

  • Prometheus + Grafana dashboards for performance and cost; health-check and load-test tooling; an admin UI for guardrail configuration and keys.

Architecture

An 11-service Docker Compose stack: the gateway, GPU inference, PostgreSQL (audit, keys, costs), Redis (cache + rate limits), the Presidio analyzer, a chat front end, admin and docs services, Prometheus, Grafana, and a reverse proxy. Self-hostable end to end — the data and the spend stay inside your boundary.

The point isn't "we use AI." It's that a team can use AI fast without giving up security, governance, or cost control — the same principle as CW DevOps, one layer down.