Local AI is most compelling when the data should not leave the machine, connectivity is unreliable, or a workflow needs an always-available model without per-request billing. The surrounding application must remain local too.
Begin with read-only material and one user. A private model connected to a public unauthenticated web interface is not a private system.
Private document and archive projects
Build household document search across manuals, warranties, insurance policies, receipts, and maintenance records. Build a family archive assistant that transcribes approved recordings, extracts names and dates, and links every answer to the original media. Build a private notebook search tool that runs on the laptop and never syncs content to a model provider.
- First version: one folder, one model, one user, visible citations.
- Success test: correct files are cited and absent answers are refused.
- Boundary: permissions and backups are explicit before indexing.
Offline field and workshop assistants
Package a small model with equipment manuals and procedures for use in a workshop, farm, vehicle, lab, or field site with poor connectivity. Build a local voice transcription and summary pipeline for notes that cannot be uploaded. Build an image-and-manual assistant that identifies a component, then points to the relevant approved procedure without pretending to replace a qualified technician.
- First version: verify the entire workflow after disabling Wi-Fi.
- Success test: source pages remain viewable beside the answer.
- Boundary: safety-critical steps require authoritative procedure and human judgment.
Always-on local services
Use a DGX Spark, Mac Studio, or RTX workstation as a trusted-LAN model server for coding, research, document processing, or experimentation. Expose an authenticated OpenAI-compatible endpoint to approved devices. Monitor memory, queue length, model version, and disk use.
Do not publish the inference or chat port directly to the internet. Use a private network, VPN, or authenticated reverse proxy with clear logs and revocation.
A local project is successful when the data path remains understandable even after the model is connected to other tools.
Primary sources