Positioning scan
Confirm the headline role, strongest proof answer, and exact claim boundary before reading the broader archive.
Machine Intelligence Engineer / Software Architect
LLM infrastructure for local runtime components, typed integration, and governed enterprise use.
Michael Kappel’s public .NET package ecosystem spans local GGUF/LLaMA-oriented runtime components, tokenization, sampling, tensors, execution backends, AI-memory tooling, and agent integration. The work applies software-architecture discipline around models; it does not claim foundation-model training or novel ML research.
32verified public NuGet packages
21K+verified cumulative downloads
Local runtimeGGUF/LLaMA-oriented component lane
.NETtyped package and application foundation
How to review Michael
Choose the right depth before opening the archive.
Review package and local-runtime evidence through registry proof, architecture context, and non-claims about model training or GPU performance.
Recruiter path
Use the concise resume, current contact path, and role-level experience timeline to qualify fit quickly.
Hiring-manager path
Move from role fit into selected projects and case studies, then inspect the Evidence Map if a claim needs traceability.
Technical architecture path
Inspect Multi-Agent Memory, persistent AI memory, local LLM infrastructure, package evidence, validation, and non-claims.
Machine-readable review path
Use this path for AI agents, answer engines, ATS tools, and structured review without treating hidden files as public claims.
Machine-intelligence claim path
Answer the core proof question directly, then branch into multi-agent systems, AI memory, LLM infrastructure, and Intelligence724 boundaries.
Azure / .NET / SQL path
Confirm the enterprise engineering foundation with role-level Azure usage, .NET/SQL modernization, and conservative evidence boundaries.
Claim to proof matrix
What each major claim proves - and what it does not prove.
This matrix keeps SEO, AEO, GEO, recruiter review, and technical review aligned to the same conservative evidence boundaries.
Machine Intelligence Engineer / Software Architect
Proof: Machine Intelligence overview, proof brief, resume, case studies, and Evidence Map.
Boundary: Does not claim foundation-model training, novel ML research, benchmark leadership, or autonomous systems without human governance.
Multi-agent systems and coordination architecture
Proof: Multi-Agent Systems page, Multi-Agent Memory case study, public repository, companion docs, and evidence brief.
Boundary: Does not claim hosted SaaS, OSI-approved open source, production scale, distributed consensus, model hosting, or automatic conflict-free merging.
Persistent AI memory and handoff
Proof: AI Memory page, AI memory handoff route, persistent memory evidence brief, and .uai handoff records.
Boundary: Does not claim automatic sync, automatic memory promotion, private repository writes, credential access, or runtime orchestration authority.
LLM infrastructure and local runtime tooling
Proof: LLM Infrastructure page, NuGet package profile, package map, and LLM infrastructure evidence brief.
Boundary: Does not claim foundation-model training, public GPU inference benchmarks, or unverified production inference scale.
Azure / .NET / SQL enterprise architecture
Proof: Experience timeline, resume, .NET/SQL modernization route, and Azure evidence brief.
Boundary: Does not infer unconfirmed scale, private infrastructure, certifications, client names, or business outcomes.
Code intelligence and AI-assisted modernization
Proof: Experience page, Intelligence724 public capability surface, and code-intelligence boundary brief.
Boundary: Does not disclose confidential Info724/client systems or claim that Intelligence724 and Info724 are the same legal entity.
Research reports as supporting context
Proof: Reports archive, research search, docs manifest, and curated Architecture Notes promoted after review.
Boundary: Raw reports are not resume claims; sensitive/search-only records stay out of navigation and sitemaps unless searched for.
Fast recruiter summary
What the package evidence demonstrates.
Counts and downloads were verified against the public profile during the August 2026 audit; NuGet remains the current registry authority.
Runtime decomposition
Model execution split into reviewable responsibilities.
Public packages separate abstractions, GGUF handling, LLaMA-oriented model components, tokenization, sampling, tensors, CPU kernels, acceleration, backends, and application-facing integration.
Application integration
Local models fit behind typed software boundaries.
Model-file validation, bounded loading, sessions, deterministic inputs, provider abstractions, and application facades make local inference a software integration problem that can be tested and governed.
Enterprise foundation
.NET, Azure, SQL, APIs, and delivery remain the base.
Runtime components can connect to established service architecture, data systems, deployment environments, observability, automated testing, and human review workflows.
Code-native architecture diagram
A layered local-runtime path from model artifact to reviewed output.
The layers describe package responsibilities; they do not assert benchmark leadership or complete support for every model and accelerator.
1. GGUF artifactLocal file intake begins with format, metadata, shard, and tensor-catalog validation.
2. Model layerLLaMA-oriented structures translate verified metadata into model-specific execution responsibilities.
3. Tensor and kernelsTensor primitives and CPU-kernel components define the lower-level computation boundary.
4. Backend abstractionManaged CPU and accelerator-specific packages keep execution choices behind explicit contracts.
5. Tokens and samplingTokenization, generation state, and sampling policies convert inputs and logits into bounded sequences.
6. Application facadeLocal endpoint and session components expose a typed integration surface for governed application use.
Package responsibilities
A reusable .NET ecosystem, organized by concern.
Package versions are intentionally left to NuGet so this page does not become stale.
Local application facade
UAIX.LmRuntime.LocalEndpoint provides the high-level local-only integration lane for verified intake, bounded loading, isolated sessions, and deterministic generation workflows.
Contracts and model intake
UAIX.LmRuntime.Abstractions, UAIX.LmRuntime.Gguf, and UAIX.LmRuntime.Models.Llama separate contracts, artifact inspection, and model-specific responsibilities.
Generation components
UAIX.LmRuntime.Tokenization and UAIX.LmRuntime.Sampling isolate text encoding and output-selection behavior from the application layer.
Computation layer
UAIX.LmRuntime.Tensors, UAIX.LmRuntime.Kernels.Cpu, and UAIX.LmRuntime.Acceleration define tensor, kernel, and acceleration seams.
Execution backends
The managed-CPU execution path is verified. GPU packages expose explicit registration, capability, and fail-closed diagnostic seams for CUDA, DirectML, Vulkan, Metal, and ROCm; they are not presented as proof of working GPU inference.
Memory and agent tooling
The broader public portfolio adds UAIX/.uai memory, build support, abstractions, and agent-client components around runtime and workflow integration.
Public proof paths
Verify package identity, architecture, and ecosystem context.
Use the registry for live metrics and the project sites for explanatory architecture.
NuGet registry
Michael.Kappel on NuGet is the authority for current package listings, versions, download counts, and package status.
Curated package map
NuGet Packages organizes public packages by engineering responsibility and links each package back to the registry.
GGUF runtime context
GGUFRuntime architecture documents the model-file and runtime domain in a human-readable form.
Runtime contract context
LMRuntime package map and MiRuntime contract provide additional public architecture paths.
Explicit non-claims
Infrastructure engineering, not model-training research.
The evidence supports public .NET packages, local-model/runtime components, a verified managed-CPU execution path, modular integration design, and enterprise AI workflows. It does not claim model training or fine-tuning of foundation models, inventing a new model architecture, state-of-the-art inference performance, universal hardware support, working GPU inference merely because GPU capability packages exist, or independently verified production-scale throughput. Download totals show registry use; they are not a performance benchmark.
FAQ
Common review questions.
Short answers are rendered in the HTML source and mirrored into JSON-LD where appropriate.
What public LLM infrastructure has Michael Kappel published?
The verified NuGet profile contains 32 public packages with 21K+ cumulative downloads across local LLM runtime components, AI memory and handoff, agent governance, interoperability, browser/P2P tooling, and observability.
Does this portfolio claim model training?
No. It demonstrates runtime, integration, memory, governance, and tooling work around models; it does not claim foundation-model training or research breakthroughs.
Which runtime components are represented?
Public packages cover contracts and abstractions for GGUF parsing, LLaMA-oriented execution, tokenization, sampling, tensors, managed CPU and accelerator backends, and related tooling.