We build AI that regulated organizations can own, from the servers it runs on to the software people use every day. Products and systems for healthcare, legal, finance, and voice operations, deployed on terms your compliance team can review.
You choose the chip, cloud, and model. We build the servers, deploy the models, and write the software around that choice. Products prove the principle. Consulting connects them to the systems, identity, and controls you already have.
We work for you, not a chip, cloud, or model vendor. We test NVIDIA and AMD hardware against the workload, and ship services as containers so you can choose where they run. How we stay vendor-neutral.
We combine open and closed models deliberately, favor local and hybrid architectures when the work demands it, and design for human review and monitoring from the start.
Co-Founder · AI & Systems Architect
Edgett has built software for organizations since 2007 and machine learning systems since Accord.NET's first release in 2010. He designs the architecture, writes the code, and when the job calls for it, builds the servers and networks it runs on. At Palm Hill he leads product and platform work across voice AI, local embeddings, document intelligence, and orchestration, and he designed Ebullience, a platform that extracts executive compensation data from SEC filings.
Before Palm Hill he spent six years at Managed Markets Insight & Technology, leading language model research, building an internal GPU cluster, and leading its contract validation product. Before that, at MFR Claims Processing, he built OCR, page classification, and image recognition into personal injury claims review, along with the data centers, network, and phone system underneath it. The full career is below.
Working with open-weight models since early 2023. Creator of Blazor Chat (opens in a new tab), contributor to LLamaSharp, HuggingSound, and more; see the open source list and the open model timeline.
Each of these comes from systems Edgett designed, shipped, and kept running.
SEC filings and executive compensation (customer work for Ebullience), drug manufacturer and insurer contracts, OCR, page classification, automatic PDF highlighting and differencing, and retrieval over messy archives. See private RAG.
Asterisk integrations going back more than a decade, a VoIP system with IBM Watson for real-time emotional and lexical call analysis, AI reception on Asterisk and FreePBX, SIP softphones, and digital audio processing.
Meta's Segment Anything models for images and video, pose estimation, Apple's Ferret (licensed for research only), Stable Diffusion, image recognition, and capture from cameras, scanners, and laser barcode readers. See vision and classic models.
Classifiers using Accord.NET since around 2010, ML.NET models that predict business behavior, custom data annotation, and small models for structured choices like Laya.
vLLM and TGI on GPUs, Ollama and GGUF locally, LLamaSharp inside .NET, LoRA fine-tuning, Whisper-class speech, and on-device embeddings for PDF and audio. Details on the Open Models page.
x86 servers and GPU clusters built for machine learning, high-performance networks with cellular modems, data centers, phone systems, and internet-connected sensors. More under hardware.
Azure, AWS, and Google Cloud, .NET Aspire and Docker services, Azure DevOps pipelines, multi-model gateways, and observability.
SQL Server 2005 through 2022 with clustering, mirroring, and CLR; Elasticsearch with vector search and Kibana; MongoDB and Snowflake; and integrations with the accounting, payroll, and payment systems a business already runs.
Private AI runs on physical machines, and someone has to pick them. Edgett designs and builds the servers and networks himself, so the person sizing the model is the same person specifying the hardware.
GPU memory decides which models fit. Power, cooling, and the network decide whether they stay fast under load. Sizing all of it against your real workload before you buy costs far less than finding out afterward. The bespoke servers we build today, on NVIDIA or AMD.
Our accessibility concept would pair camera glasses and a phone with a tactile braille cuff. Early concept. Not a product. Illustrations only.
Most of it for organizations where the records have to be right: health data, legal claims, payroll, and accounting.
If your system isn't listed, ask. A lot of this list came from connecting to software someone else chose.
| Area | Tools |
|---|---|
| Languages | C#, Python, SQL, JavaScript and TypeScript, Java, C and C++, Dart, Scala, XAML |
| Interfaces | Blazor, Angular, React, WPF, .NET MAUI, Flutter, ASP.NET MVC and Web Forms, Streamlit |
| Data and search | SQL Server, Elasticsearch and Kibana, MongoDB, Cosmos DB, PostgreSQL, MySQL, Snowflake, Databricks |
| Machine learning | PyTorch, Hugging Face, ONNX Runtime, ML.NET, Accord.NET, Azure AI and Document Intelligence, AWS Bedrock and SageMaker, IBM Watson |
| Voice | Asterisk, FreePBX, SIP, Whisper, wav2vec2, Deepgram |
| Cloud and DevOps | Azure, AWS, Google Cloud, IBM Cloud, Docker, .NET Aspire, Azure DevOps, GitHub |
| Infrastructure | vSphere, Hyper-V, Proxmox, SQL Server clustering and mirroring, backup and disaster recovery |
| Hardware | x86 GPU servers with NVIDIA and AMD GPUs, GPU clusters, high-performance networks, cellular modems, Raspberry Pi, Arduino, IoT sensors |
| Platforms | Windows, macOS, Linux, Android, iOS |
Projects Edgett created and merged pull requests to projects he uses.
PalmHill.BlazorChat (opens in a new tab): chat with local LLMs on .NET 8, with document retrieval. Created November 2023.
PR #386 (opens in a new tab) fixed the AVX512 Windows build for the .NET llama.cpp bindings. Merged December 2023.
PR #92 (opens in a new tab) restored installs on Torch 2 and Python 3.11. Merged August 2023.
PR #59 (opens in a new tab) fixed how paragraphs are modeled in transcription results. Merged January 2023.
Azure DevOps plugin (opens in a new tab) for searching code and work items from the Windows launcher. Created February 2023.
PR #12035 (opens in a new tab) corrected the Scoop install command in the docs. Merged April 2022.
PR #2355 (opens in a new tab) added a CloseOnClickOutside option to the Popup component. Merged November 2025.
PR #26 (opens in a new tab) made the workspace folder configurable so document conversion works in read-only environments such as serverless hosts. Merged February 2026.
Whether you need a product deployment, a custom system, or the servers to run it on, start with a direct conversation.
