AI & LLM Infrastructure — The Free-Tier Architect

Anyone can call a paid API. The interesting engineering is building infrastructure on free tiers that behaves like a paid one — and that is precisely what my AI projects do, in layers.

The layers

Layer	Project	Job
Access	GLM Free API, GhostBrain, qwen/deepseek bridges	Turn free web models into real OpenAI/Anthropic-compatible APIs.
Routing	OmniRouter	Load balancing, weights, quotas, failover across providers.
Coordination	Hermes Stack	Agent runtime + automations + console, self-hosted and free.
Consumption	OpenCode · Codex · Claude Code	Agent fleet doing real engineering work through the stack.

The design rules are consistent: never depend on one provider, never trust a free tier to stay free, always keep an escape hatch. When a bridge dies, the router benches it; when a platform turns hostile, the panel discipline applies — migrate in minutes. The result is a multi-model AI stack with a monthly bill of exactly zero.

On the ML side itself I work with TensorFlow, PyTorch, OpenCV and Hugging Face tooling — enough depth to fine-tune, embed and evaluate, which is usually all a product actually needs.