Llama, Qwen, Mistral, Gemma, and Whisper-class models, plus vision and classic machine learning models, served on hardware or private cloud you control. We have worked with open-weight models since early 2023 and use them in our products today.
Serving options, fine-tuning, and speech models. We choose them after testing your workload.
Quantized models on CPUs and workstation GPUs. Suited to local assistants, edge boxes, and air-gapped sites.
High-throughput GPU serving with continuous batching and OpenAI-compatible endpoints for team-scale workloads.
In-process inference inside C# services and desktop apps. No Python sidecar and no separate model server to operate.
Open-weight models hosted in your own cloud account when you want managed GPUs with your existing identity and billing.
Adapters trained with transformers and PEFT on your data, evaluated against your tasks before they reach production.
Transcription and audio understanding that stays on your machines, for call recordings, dictation, and archives.
With a local deployment, model requests stay on your infrastructure. You control the hardware budget and model version. If you choose a hosted model, we document which requests go to that provider.
Not every job suits an open model. We say so in discovery and bring in closed models where that is the better answer.
Laya is a small open model we run for structured choices: pick from your list, score on your scale, or answer yes or no when a large language model is the wrong tool. It can run on your hardware, with people reviewing consequential decisions.
For example, a model could suggest which team should handle a support ticket. We test its suggestions on your examples and decide with you when a person must review them. A fixed set of answers does not guarantee a correct choice.
Chat411 is part of PalmConnect, an early release. It answers and transfers calls on Asterisk or FreePBX and supports two conversation paths today.
Chat411 can use a local model on this same stack. Media can be encrypted, and PSTN access is not required. Internal extensions, intercom, on-net SIP, and closed voice networks can keep conversation audio on infrastructure you control without a cloud model provider.
Chat411 can also stream caller audio to a cloud realtime voice provider, xAI by default, when you accept that egress. Call control stays on your PBX. This is a separate choice from the local model path.
Plenty of useful models never write a sentence. Some look at images and video. Others are small classifiers trained on your own records, and they are often the fastest and cheapest answer.
Meta's Segment Anything (opens in a new tab) models outline any object in an image or video from a click or a box. They are released under Apache 2.0, run on your own GPUs, and make a good first step for measuring, counting, or redacting what is in the frame.
Pose models find the joints of a body in each frame, turning video into coordinates your software can check for posture and movement. They run on local GPUs, so the footage stays with you.
Apple's Ferret (opens in a new tab) answers questions about a region you point to and marks the objects it names. Apple licenses it for research use only, so it suits evaluation and prototypes; production work moves to a model with commercial terms.
Edgett has worked with the Accord.NET Framework (opens in a new tab) since its first public release in May 2010, when it brought support vector machines, decision trees, and computer vision to C#. Later work moved to ML.NET, including custom models that predict business behavior, and to page classification and image recognition for claims documents.
A classifier trained on your own records can run on an ordinary CPU and be retrained quickly when the data shifts. When the question is narrow and you have labeled examples, it is often the right tool. Models such as Laya can return a choice or score from text when your application needs a structured answer.
The project was archived in November 2020, so it no longer gets fixes. We can keep your models running as they are, or retrain them on a maintained library such as ML.NET and check that the new model agrees with the old one before you switch.
Two of our products run open-weight embedding models on the user's own device.
PDF Diff aligns pages between documents with EmbeddingGemma (opens in a new tab) (google/embeddinggemma-300m) running locally on Mac and Windows. Documents never leave the computer. Free download.
Audio Search runs multimodal text and audio embeddings on-device with ONNX Runtime, so text-to-audio and audio-to-audio search work without uploading a file. Available by request.
The same on-device embeddings are the retrieval half of the private RAG systems we build.
Small, merged, and linked. Each one came from running these tools in real work.
PalmHill.BlazorChat (opens in a new tab): local LLM chat and API on Blazor WebAssembly, SignalR, and .NET 8, with Llama 2 and Orca 2 support. Created November 2023.
PR #92 (opens in a new tab) restored installs on Torch 2 and Python 3.11 for a speech toolkit built on Hugging Face Transformers. Merged August 2023.
PR #386 (opens in a new tab) fixed the AVX512 Windows native build for the .NET llama.cpp bindings. Merged December 2023.
Each entry sits next to the release it followed, so the gap is visible.
Shortlist candidate models, review licenses, size the hardware, and stand up a serving path. Evaluate on your data and leave with numbers, not opinions.
Gateway, auth, observability, fine-tuning where it pays off, and runbooks for upgrades and rollback so your team owns the deployment.
Schedule a discovery session. We will map open-weight options to your hardware, data boundaries, and workload.
