A local-first AI vision experiment

Give a local AI
something to see.

Jase Vision AI is a browser-based vision assistant powered by MiniCPM-V running locally through Ollama. Visitors can explicitly share a webcam frame or upload an image and receive a natural-language description from the locally hosted model.

MiniCPM-VOllamaPython / FlaskBrowser webcamImage uploadCloudflare Access

What it can do

Live webcam capture

The visitor chooses to start their own browser camera and explicitly grants permission before a frame can be captured.

Image uploads

Accepts JPG, PNG and WEBP files, with an 8 MB per-image limit and basic file-signature checks.

Local model inference

Flask forwards image bytes to Ollama on the host computer, where MiniCPM-V generates a concise description.

Invite-only access

The hosted demonstration sits behind Cloudflare Access. The public repository does not grant access to the running service.

How the request flows

Visitor browser ├─ webcam permission and frame capture └─ image file selection │ HTTPS ▼ Cloudflare Access (invite-only) │ ▼ Cloudflare Tunnel │ ▼ Flask application (127.0.0.1) │ ▼ Ollama local API │ ▼ MiniCPM-V vision model │ ▼ Description returned to the visitor's request

Built with

Python + Flask

Small web backend with separate webcam and multipart upload endpoints.

Ollama

Local model runtime serving the vision-language model through its local API.

MiniCPM-V

A compact vision-language model used to interpret the submitted image.

Cloudflare

Optional HTTPS tunnel and identity-aware access layer for controlled remote testing.

Privacy and current status

Invite-only demo. The live service is not open for unrestricted public use. Access is controlled separately from this public source repository.

The Flask application processes image bytes in memory and does not save uploads to a permanent application folder. The browser sends the image to the hosted application for inference. Other components such as the browser, operating system, tunnel provider and model runtime may process technical metadata under their own policies.

This is a working personal project and controlled demonstration, not a hardened multi-tenant production service. Rate limiting, per-user quotas, malware scanning, a job queue and concurrency scheduling remain future improvements.