HomeAI Systems Lab

Forward-Deployed AI Solutions / evidence-bearing systems

AI-directed engineering with human acceptance authority.

The primary research direction is local geometric and latent reasoning. The lab also isolates my implemented AI systems from the conventional hardware, infrastructure, and software portfolio, showing how I translate operating problems into bounded contracts, direct specialist agents, challenge assumptions, and require evidence before accepting a result.

Primary research: Geometric / latent reasoningResearch system: NVIDIA DGX Spark with GB10Positioning: Forward-Deployed AI SolutionsTarget role: Forward-Deployed AI EngineerCurrent formal title: Technician III

R1 / primary research direction

Geometric reasoning under controlled local evaluation.

I direct a DGX Spark research program testing whether latent or compressed reasoning can reduce model work without sacrificing correctness, explanation integrity, or useful general behavior.

Research question

Efficiency that preserves quality

Translate GLR and related latent-reasoning ideas into falsifiable experiments that measure correctness, stopping, tokens, latency, and explanation quality together.

Controlled evidence

Matched training and evaluation

Completed frozen-baseline comparisons, prompt and stopping diagnostics, direct-answer controls, and a task-mix audit on the local GB10 system.

Current decision

Redesign before scaling

The tested adaptations became shorter or cleaner without preserving answer quality. Further training is held while the task mix and reasoning targets are redesigned.

Open the research record

AI-directed engineering method

The model proposes. The system evidence decides.

I use AI agents as specialized engineering collaborators while retaining ownership of requirements, authority boundaries, tradeoffs, correction, and final acceptance.

01
Discover the actual field problem
Observe the work, separate symptoms from causes, and identify what meaningful outcome the operator needs.
02
Map constraints and authority
Define data, identity, security, lifecycle, privacy, resource, and human-decision boundaries before implementation.
03
Write contracts and acceptance gates
Turn observations into versioned interfaces, failure behavior, invariants, and tests that can reject a plausible but wrong result.
04
Direct bounded workstreams
Assign modular specialist work, preserve ownership boundaries, and review architecture and tradeoffs instead of treating generated code as finished work.
05
Test at the right evidence layer
Distinguish source review, unit checks, integration harnesses, real executables, installation, network behavior, and live operating evidence.
06
Preserve uncertainty and hand off
Keep known failures and pending acceptance visible; package, document, and transfer a result that another engineer can understand and operate.

Featured case studies

Seven systems at clearly stated maturity levels.

Each page shows what Shayne owned, what AI contributed, where the system was exercised, and the next meaningful validation stage. Metrics and test counts are presented at the engineering layer where they were measured.

Directed & checksum-verified

Self-hosted AI infrastructure · storage operations

AI model-storage migration

Directed and acceptance-tested a nearly 600 GB model migration, preserving integrity across hundreds of files while recovering substantial system capacity and creating more than a terabyte of growth headroom.

Read the evidence

Installed & live-verified

Windows agent execution · .NET 10

Pegasus AgentOps

A typed, auditable execution subsystem shared by MCP and a product-local CLI, packaged and exercised through real install and repair on one developer workstation.

Read the evidence

Prototype & tested

Human-in-the-loop product evaluation

Neural Commander

A deterministic Godot/.NET RTS vertical slice shaped by playtesting, explicit simulation contracts, content admission, and native visual/audio evidence—with one known deterministic test failure still tracked.

Read the evidence

Two-computer verified

Cross-computer AI coordination

Pegasus Agent Mail

Bounded file-carried context with metadata-only wake notification, exact task identity, exclusive pickup, repeat-request reuse, and clean-queue acceptance between two Windows computers.

Read the evidence

Current correction pending

Local-first relay · delivery lifecycle

Pegasus Exchange

A substantial peer-node, Desktop, CLI, installer, and bridge system with earlier two-machine proof. Its newest scheduling correction remains a candidate awaiting renewed acceptance.

Read the evidence

Private alpha & tested

Privacy · deterministic explanation

Personality Synthesis Engine

An accountless, client-held self-reflection application with governed content, deterministic routing, strong test coverage, and explicit readiness gates for its next release stage.

Read the evidence

Historical project record

Offline developer tools · semantic visualization

CodeConstellation

Directed product requirements and UAT for an offline semantic 3D code map spanning a substantial object and relationship graph, with camera preservation and an Obsidian continuation path.

Read the historical evidence

Secondary architecture record

Pegasus Coder: approval-controlled AI development.

A Codex-like platform architecture for durable jobs, checkpoints, audit, recovery, and fail-closed local execution. The current record supports product direction, authority design, and historical focused checks; provider integration and end-to-end runtime validation are the next milestones.

Supported contributionProduct ownership, requirements authority, approval and recovery boundaries, and architecture direction.

Recorded project evidenceA multi-project architecture with historical release and focused-check records preserved in the curated project package.

Authority modelModels and prompts cannot grant themselves execution authority; admission, audit, leases, and recovery remain explicit system responsibilities.

Current maturityArchitecture and historical checks, progressing toward provider integration and end-to-end runtime acceptance.

For live conventional web infrastructure, see the Professional Portfolio

Maturity language

Every stage identifies what exists now.

Research, design, implementation, testing, packaging, installation, and live verification describe progressively stronger evidence. Each case study names its present stage and the scope in which it was exercised.

Research / DesignedA defined question, architecture, or product direction with a concrete evaluation path.

Prototype / ImplementedWorking behavior in a bounded environment, ready for progressively broader validation.

Tested / PackagedNamed checks and reproducible release artifacts at the stated scope.

Installed / Live-verifiedA real runtime or multi-machine scenario exercised in the environment described by the case study.

Technology and capability matrix

Grouped by engineering responsibility.

Languages are listed as used technologies without invented proficiency ratings. Product-specific frameworks appear only where the case-study evidence supports them.

Systems integration

Field discovery and workflow mapping
Root-cause analysis and failure paths
Hardware, software, network, and operator boundaries
Supportable handoff and recovery