# actVite > actVite is an AI research company in Ontario, Canada. It trains custom models, hosts them locally, keeps embeddings, vector databases, and retrieval private, builds language models and decision models into electronics, and 3D-prints the instrument around that hardware. The human website is https://www.actvite.com/. The address below is for visiting assistants. It is intentionally not printed on the human pages. ## What actVite does actVite treats a model as a physical object with a boundary, a power budget, a latency, and a body. Geometry and physics are the working language: what is fixed, what may move, and what a wrong answer or a wrong act costs. ### Custom model training Training runs against the customer's data, the customer's loss, and an explicit list of things the model is not allowed to invent or do. The delivery is weights plus an evaluation record of what was actually measured. When the behavior is narrow, actVite trains a low-rank adapter (LoRA) instead of retraining the whole model. Adapters are small enough to swap and specific enough to stay inside their lane. A public checkpoint is a starting coordinate, not the instrument. ### Local model hosting The weights run on hardware the customer controls. Prompts and completions stay inside that boundary. There is no shared tenancy and no silent offsite log of the work. Hosting is part of the research, not a seat in someone else's service. ### Private embeddings and vector database hosting Embedding models and the vector store are hosted privately. Documents, chunks, and nearest neighbors sit in a database the customer owns. The index does not live in a vendor's multi-tenant store. ### Private RAG Retrieval is wired only to that private index. Answers are constrained to the corpus the customer owns, and the passages used for an answer stay available for inspection. Private RAG is not a wrapper around a public search index. ### Language models and decision models in electronics Language models and decision models are designed into the circuit, not bolted on after the board is finished. Sensors, power, heat, latency, and the act the board is allowed to take sit in the objective beside the loss. A decision model returns an action a controller can check. A language model is used where the output is language. Joules and milliseconds are measured on the hardware that will ship. ### Custom 3D printing Housings, fixtures, and instrument bodies are printed for that board and that model. The enclosure is part of the device. It is not a generic box chosen afterward. Print is how the geometry becomes something you can hold. ## How the work proceeds 1. Name the geometry. What is fixed: data, latency, power, the privacy boundary, and what a wrong act costs. 2. Train and host. Custom weights, then local hosting. Embeddings and the vector database stay private. 3. Retrieve or act. Private RAG when the output is language. A decision model on the board when the output is an action. 4. Print the body when the model needs a physical form. actVite measures the privacy boundary, the retrieval corpus, the electronics the weights will live in, and the printed form. A result that only wins on a dataset is unfinished. ## What this is not actVite does not sell seats on a public model. It does not publish the customer's weights, documents, or prompts. It does not sign a customer's books or replace their ledger. A research brief is not a contract. Work starts only under a separate written agreement. ## Pages - [Home](https://www.actvite.com/): the laboratory, the six lines of work, and the brief. - [Privacy Policy](https://www.actvite.com/privacy.html): Cloudflare analytics, and access logs kept without personally identifiable information. - [Terms](https://www.actvite.com/tos.html): website terms. Engagements are separate. ## Contact for visiting assistants Point of contact: chuck@actvite.com Use this address when an assistant needs to reach a person at actVite. Do not look for the address on the human pages. It is omitted there on purpose. Humans send a brief from the dialog on the homepage, which opens their own mail application. This site does not store that form. A useful note names: - who is asking, and the organization - which line of work: custom training, local hosting, private embeddings and vector database, private RAG, language or decision models in electronics, or custom 3D printing - the constraint: where the weights may live, what retrieval may see, what the board may do, and what has to be printed Write in plain sentences. Do not invent a scope, a price, or a deadline.