Work
Case studies, not source
Most of this work is private and belongs to the organizations it serves, so each piece names the problem, what was built or decided, how it was governed, and the honest outcome, including the ones where the right answer was to stop.
Story Maker: a governed AI-agent platform
The decision that mattered: no agent runs unattended until a person has approved the exact list of operations it may perform.
Technicians build and run AI agents against real business systems under tiered least-privilege skills. A human in the loop approves every operation before the agent runs alone; a supervising agent halts anything outside the approved plan. Five inference tiers under U.S.-residency rules; every call attributed for cost. 3,277 automated tests.
Architecture and security posture ->County AI and AI-governance roadmap
The decision that mattered: write the roadmap with the county's own IT team, not for them.
Working alongside a county IT and network team on systems, network, and security operations while helping them write an AI roadmap they can follow: where sensitive data is exposed today, tiered guardrails, in-house inference with controls and visibility, and the staff-augmentation uses worth doing first.
Ask about this engagement ->The Catalog: 260 work tools scored for AI-friendliness
The decision that mattered: publish the rubric. A score you cannot inspect is just another opinion.
Every tool in the catalog gets a 0-100 AI-friendliness score from the same seven-part rubric, covering whether the software will connect to an AI working on your behalf, hand over its data, and do so safely. Tiers: AI-Native, AI-Ready, AI-Limited. A stack check scores the tools an organization already pays for in about a minute, with no sign-up. Published scores are verified against vendor documentation and re-reviewed at least every six months. No vendor pays for a score.
Open the catalog ->Melon Code: a Claude-Code-class agent on a local model
The decision that mattered: run the whole coding-agent loop on a model we own, and point Claude Code at it to learn how the loop really works.
The full coding-agent loop, with tools, a four-mode permission engine, subagents, compaction, and rewind, driving a 284-billion-parameter model on one workstation. An Anthropic-compatible endpoint lets Claude Code itself run on local or hosted open models. KV-cache prefix work cut a warm turn from 36 seconds to 1.3.
What was built and measured ->Local inference at the edge of one GPU
The decision that mattered: measure before optimizing. Three speculative-decoding routes closed with break-even math; one small change kept.
Decode raised from 7.3 to about 20 tokens per second at a stated fidelity floor. A DSpark confidence-scheduled drafter, a multi-token-prediction head, and DFlash each closed with numbers; the DSpark paper's prefix-survival gate implemented in llama.cpp for a measured 7.1% gain. 190+ journal entries, pre-registered decisions.
The research list ->A phone-sized model, built from nothing on one desk
The decision that mattered: start from random weights and a 27B teacher, pre-register everything, publish the journal.
A mixture-of-experts language model in the Qwen3.8-Flash-Next architecture family, sized to the iPhone Air's 2.5 GB safe memory tier, taught by distillation from Qwen3.8 27B, with DSpark and MTP speculation measured on the phone with reasoning on.
Follow the journal ->MelonStudio: splitting a model across GPU and CPU by hand
The decision that mattered: stop. A negative result measured honestly, and the reason later work moved to llama.cpp.
ONNX Runtime won't offload layers the way llama.cpp does, so this built it: graph partitioning at layer boundaries, two orchestrated sessions, pinned memory, a stateful KV cache. Measured at 4.4 tokens per second hybrid against 93 GPU-only, the result was to stop.
Design, benchmarks, and the negative result ->Engineering with people and agents
The decision that mattered: write the operating model down so people and agents work the same way.
Audit-first work packets, isolated worktrees, worker, verifier, and adversarial-reviewer roles, a test gate on every pull request, and a reservation ledger so concurrent agent sessions share one machine without collisions.
The process, and the ledger's source ->