Patrick Reynolds · Guy Prompting

Put AI to work for your people, with limits and someone watching.

People matter.

Technology exists to take the grinding work off people so they can do the work only people can do. The aim is never a smaller team. It is the team you already have, twice as effective.

Applications must be AI friendly.

An organization adopts AI well when its systems expose real, governed interfaces, its identities carry least privilege, and its data handling is written down. Most of what goes wrong with AI is an access problem.

How a governed run works · try it

Give an agent a task. Watch what it is allowed to do.

This is the pattern behind the platform: skills are tiered, a person approves the plan before the agent runs alone, and a supervising agent halts anything off-plan. Pick a task and a tier, then run it.

The run log appears here.

Rehearsal firstThe agent walks the plan with zero live calls and lists every operation it intends. A person approves that list once.

Closed worldUnattended runs may only perform approved operations, reads included. Anything else is blocked or parked.

Director on setA supervising agent screens intent on the actual arguments, not just the tool name, and halts deviations.

Every call has a nameInference is attributed to an agent, so cost and behavior are always traceable.

Selected work

A phone-sized model, built from nothing on one desk

In progressSept 2026Qwen 4-class MoE · iPhone Air

A mixture-of-experts language model with the Qwen3.8-Flash-Next architecture family (Gated DeltaNet and sparse attention, gated residual, an n-gram table, a multi-token-prediction head), sized to the iPhone Air's 2.5 GB safe memory tier, started from random weights, taught by distillation from Qwen3.8 27B, with DSpark and MTP speculation measured on the phone with reasoning on. Every decision pre-registered; the journal is the deliverable.

Follow the journal →

Governed AI-agent platform

Production202633 client organizations

Technicians build and run AI agents against real business systems under tiered least-privilege skills. A person approves every operation before the agent runs unattended; 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 →

Melon Code: a Claude-Code-class agent on a local model

Daily use2026TypeScript · llama.cpp · NVIDIA NIM

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

Research2026284B mixture-of-experts · RTX 4090

Decode raised from 7.3 to about 20 tokens per second at a stated fidelity floor. Three speculative-decoding routes (a DSpark confidence-scheduled drafter, a multi-token-prediction head, DFlash) each closed with numbers from derived break-even math; 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 →

MelonStudio: splitting a model across GPU and CPU by hand

Complete · negative result2025–26C# · ONNX Runtime GenAI

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 honestly (4.4 tokens per second hybrid against 93 GPU-only), the result was to stop and move to llama.cpp.

Design, benchmarks, and the negative result →

County AI and AI-governance roadmap

In progress2026Georgia county government

Working alongside a county IT and network team on systems, network, and security operations while helping them write an AI roadmap they can actually follow: where sensitive data is exposed today, tiered guardrails, and the staff-augmentation uses worth doing first.

Ask about this engagement →

Engineering with people and agents

In daily use2026process + ledger

The written operating model behind the platform: 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 →

Live demonstrations

Live, sign-in-gated demonstrations are being added here: a local language model running on the workstation under my desk, the governed agent platform with a fictional client set, and the endpoint control plane driving a simulated fleet. Each runs as a separate instance with synthetic data, the safe-read tier by default, nightly reset, and an identity check in front. The demos are governed the same way the real systems are, because that is the point.

Research, with the negative results kept

Shepherd: a self-hosted endpoint control plane and Windows agent with tenant isolation, signed tasking, and a hash-chained audit log, built so an AI agent can operate a fleet safely. Working prototype, ten decision records.

Local inference of a 284-billion-parameter mixture-of-experts model on one workstation: decode raised from 7.3 to about 20 tokens per second; 190+ numbered journal entries and pre-registered decisions.

Lossless LLM-driven text compression: 0.882 bits per byte on enwik8 against 7-Zip's 1.989, with a measured ratio-versus-speed frontier.

A deterministic context manager for long agent sessions: byte-exact log, lean resident window, 93–100% recall of exact facts at 90–99% token reduction, claim narrowed after an adversarial audit.

On-device model selection for phones: an MMLU-Pro harness projected to memory bandwidth across 298 candidate models.

A self-hosted, Claude-Code-class coding studio with an Anthropic-compatible endpoint, so standard agent tooling runs on local or cloud models.

A de-identified calibration corpus built from real platform traffic, with two adversarial privacy audits; the first said "do not ship."

About

I'm Patrick Reynolds. Since 2004 I have run Cross Link Consulting, a security-first IT firm of about fifteen people serving county and municipal governments, doctors, lawyers, accounting firms, and industrial clients across the Augusta area. We have provided breach monitoring and response since 2019, finding and evicting intruders before data leaves.

I founded Resolute AIM to help organizations stop asking AI questions and start putting AI to work: guidance, talks, governance, and, when it is needed, the application itself. For the last year I have spent most of my working hours building AI platforms with my own hands, and the lesson is consistent. AI adoption succeeds on identity, access, and data handling long before it succeeds on models.

Before any of that: undergraduate study at USC Aiken, technology and electronics training in the U.S. Army, eight years in the Reserve, five years keeping a hospital's clinical equipment running, then the Microsoft certifications that opened the door to systems administration. I hire for character and teach expertise. Serving humbly. Caring well.