I build , AI systems for when they can't be wrong.
StoneyTECH · AI Architecture Applied · authored by Auston DeVille
Mostly that comes down to one question, asked over and over: how much should I let the model decide here, and how much should I take off its plate and hand to something I can audit? This site works that question in the open through cited essays, reference builds, and the rules that stay useful under real constraints.
Some of what follows is written for you. Some of it is written for the agent reading on your behalf, through the read-only MCP. It's the same material either way, because if I can't hand a machine a clean version of my thinking, I haven't finished thinking it.
Start Here
Three ways in: learn the vocabulary, start with the Determinism Ladder, or point an agent at the bounded public MCP.
- New to AI systems?
Build a working vocabulary for , , , and .
- Designing production AI?
Start with the : where should responsibility live when model autonomy gets risky?
- Pointing an agent here?
Use the , the read-only interface for published StoneyTECH context.
Latest from Learn
- Everything Gets Rebuilt Hugging Face's responders reached for their frontier models mid-breach and the models refused the evidence. Accurate, intentional, secure AI comes before the fire, not during it.
- The Builder Goes First Why frontier AI tilts the attacker-defender fight toward whoever builds first, and why the optimism comes with a body count.
- The Discipline Patch A small corrective fine-tune turns failed evals into behavior training: patch the discipline, not the facts.
What this is
A public notebook on the : citation-first, public-source traces, small reference rebuilds, and a running record of what held and what broke. The posture stays grounded: I cite stronger work, restate it in practical language, and show the proof path behind each claim.
How the site judges itself
The mission does not just describe bounded systems. The site submits itself to the same test: public identity boundary, receipt-backed claims, read-only agent surface, and now a public graph agents can traverse directly.
Disclosure
StoneyTECH.net is independent learning and portfolio work. It is not affiliated with, endorsed by, or representative of any employer. The public material here does not use employer confidential information, employer systems, unpublished work product, proprietary customer data, or private original work. It makes no claim to original research; anything genuinely original stays out of the public corpus.
- Demystify AI — plain-language primers for overloaded AI terms.
- Learn — essays on the LLM & agentic stack, from LoRA to LangGraph. Each piece tests the metaphor, the trade-off, and the public reference trail.
- MCP — the read-only agent interface for published StoneyTECH context.
- Builds — walkthroughs of the systems behind the essays.
- About — site posture, boundaries, and learning method.
- Monitor — public observation notes and published status surfaces.