AI infrastructure / homelab

Local AI infrastructure as technical capability.

This is not a gaming build page. It is an operations-minded environment for learning inference tradeoffs, deployment patterns, security boundaries, and practical AI service design.

Local inference

Model-family testing across Qwen 3.5, Qwen 3.6, Gemma 4, GLM 4.7 Flash, and Bonsai ternary models for understanding capability, latency, memory pressure, and task fit outside hosted APIs.

Hardware constraints

RTX 4060 Ti 16GB class testing for VRAM, context, throughput, and quantisation tradeoffs.

Mini-server stack

Docker, Caddy, vLLM, llama.cpp, LM Studio, LangChain, PostgreSQL, pgvector, pgai, ChromaDB, FAISS, reverse proxying, and public demo hosting patterns.

Security-conscious design

Local inference, restricted endpoints, public frontends, protected backends, and careful exposure boundaries.

RAG deployment

Retrieval pipelines, reranking, answer grounding, prompt-injection protection, guardrails, source constraints, and operational monitoring considerations.

Deployment constraints

The same constraints appear repeatedly: cost, risk, throughput, data boundaries, and maintainability.