Isolated execution
The sandbox isolates agent work with controlled access to repositories, tooling, and credentials. It routes hosted and local models into bounded, repeatable tasks so I can replace tooling without destabilizing my workstation.
AI Lab / my personal workshop
The lab exists to compress the engineering feedback loop. I build and operate software and AI systems daily in my own environment, moving from ideas to evaluation and revision. Current goals include faster reasoning, lower token use, and more efficient agent communication; model quality and reliable results remain constraints. That experience informs my enterprise architecture work, where security, governance, reliability, and team ownership serve a different scale.
AI Lab / environment
Much of this environment is in daily use. I build the surrounding infrastructure as well as the AI workflows; nearly all AI-assisted development runs in a dedicated server-side sandbox.
The sandbox isolates agent work with controlled access to repositories, tooling, and credentials. It routes hosted and local models into bounded, repeatable tasks so I can replace tooling without destabilizing my workstation.
One AMD Strix Halo machine with 128 GB unified memory supports local inference and workers. I benchmark context size, quantization, runtime, task quality, and routing against hosted options. The practical question is which useful work smaller local systems can handle reliably.
I run VMs, databases, analytics, monitoring, storage and backups alongside applications and services. Server-hosted Supabase supports structured database and vector storage for AI workflows. Deployment and recovery belong to the same operating stack.
Orchestration, validation, and review give bounded agent work a path to inspectable results. Model evaluations guide routing and workflow improvements. That evidence may inform future fine-tuning and specialist experiments; LoRA and CPU-sized models remain exploratory.
AI Lab / operating loop
I do this because architecture, infrastructure, AI, developer tooling, and product engineering are the parts of engineering I most enjoy—and this environment lets me work across all of them.
Idea → architecture → implementation → deployment → real use → evidence → opinion → revision. Independent decisions, hands-on operations, and rapid replacement make that loop useful in my own environment.
I start with a small deterministic tool where it is enough, use a capable local model where it holds up under validation, and call on frontier reasoning when the problem warrants it. The cards below show which work is built and used, being benchmarked or tested, being built, or still a future investigation.
AI Lab / systems and experiments
Built work, benchmarks, active tests, and future investigations each have their own stage. Results from real use inform the next iteration and my professional judgment.
I benchmark local inference on my AMD Strix Halo machine with 128 GB unified memory, comparing context, quantization, runtimes, latency, and task quality. Faster models and reasoning are optimization goals; I measure tradeoffs before changing routing.
I built and use indexes, syntax-aware analysis, repository graphs, and targeted retrieval to find relevant code before model reasoning.
I am testing task-scoped context and concise reasoning strategies aimed at reducing input, output, and reasoning tokens while preserving quality. Lower token use and faster reasoning are design goals, not claimed measured gains.
I use repeatable tasks and evaluator feedback to compare model and workflow performance, then improve routing and agent workflows based on the results.
I am exploring automated independent AI review of code, validation evidence, and rendered website changes, including visual UX checks. The goal is to return specific findings for bounded correction, not open-ended redesign.
I am exploring structured, more compact ways for agents to exchange instructions and results. The goal is faster coordination with fewer communication tokens, without sacrificing clarity or task quality.
I am building long-lived, dependency-aware pipelines that coordinate bounded subagents, evaluation, UX review, deterministic checks, retries, and approval. Fewer tries and less back-and-forth are design goals, not guaranteed outcomes.
I structure interaction records and outcomes for performance comparisons, failure analysis, and workflow improvements. Evaluation evidence can also inform future fine-tuning experiments.
I use Supabase-backed database and vector storage to support retrieval for AI workflows, with attention to source provenance, service ownership, and relevance.
I am exploring LoRA and small specialist models for narrow tasks such as classification, bounded decisions, completion checks, and a CPU-friendly game companion. These are experiments, not production deployments.