A 4-GPU home lab on the Mississippi Gulf Coast runs Qwen, Gemma, and a rotating set of open-weight models. Each machine has one job, and a small custom gateway called homelab-router decides which one answers each request. When iterating on multi-agent architectures and context engineering, fast local feedback loops matter more than raw peak throughput: no cloud costs, no round-trip latency, no queue for a shared GPU.
Lab · Stack
The operating stack behind the work.
The tools, frameworks, and hardware used to build agentic AI systems, with emphasis on local control, observable operations, and open-source foundations.
Home Lab Philosophy
Home GPU Lab
Machine names and VRAM are derived directly from the authoritative HomeLab dataset. Descriptions map to each machine's assigned job.
- ubuntu-6000RTX PRO 6000 Blackwell (96 GB VRAM)Heavy code generation, refactoring, agentic / multi-file work.
- SparkNVIDIA DGX Spark (Unified CPU + GPU memory)Senior reasoning, architecture, final review, safety fallback.
- win-5090RTX 5090 (32 GB VRAM)General chat, rewrites, medium-context summarization, spillover.
- ubuntu-4090RTX 4090 Laptop GPU (16 GB VRAM)Fast vision, JSON utility, low-latency routing pre-filter.
Development Environment
AI Coding Assistants
Cloud & Infrastructure
AI/ML Platforms
RAG & Agent Frameworks
Vector Databases
Python & Deep Learning
Frontend & Web
Currently Experimenting With
Concrete experiments currently underway across production, research, prototyping, and exploration tracks.
- 12-15 Agent RAG PipelinesSpecialized agents for retrieval optimization at terabyte scale.
- Context Engineering TechniquesMemory optimization to overcome attention dilution.
- Plan-Act-Learn PipelinesSelf-learning ETL that adapts autonomously.
- Real-Time Hallucination DetectionQuality feedback loops with continuous learning.
- Local Fine-Tuning WorkflowsLoRA and QLoRA on consumer GPUs for domain adaptation.
- Headless SaaS PatternsAgent-first API design for the post-GUI era.
Research Focus Areas
- Multi-agent parallel architectures for retrieval optimization
- Recency bias and attention dilution mitigation
- Reinforcement learning for RAG response quality
- Forward-thinking entity-linking for feature spaces
- Autonomous data pipeline adaptation
- Agent-to-agent communication protocols
Open Source & Community
Most of the stack sits on open frameworks — LlamaIndex, LangChain, Semantic Kernel, Weaviate, pgvector. These communities move faster than any single vendor. When the ideas are flowing, the work is code: tweaking RAG pipelines, fine-tuning models on local GPUs, and wiring the results back into agentic workflows.