Can I trust a green border as proof a post is safe?

No. Percentages are model signals, and Instagram Ad, Sponsored, and Paid partnership labels override the models with an orange border. Treat green and blue as hints, not verdicts.

What happens when Gemini is down but Jev still works?

Yes. The backend combines both signal sets, explains which provider led, and falls back visibly to text-only results when visual analysis is unavailable.

What data leaves my browser, and are frames stored?

Only visible text, alt text, content type, surface, and page language go to TypeSafe, plus downscaled images or up to three Reel frames to Gemini. Frames stay in memory and are not saved.

How do I run it without it touching my Instagram account?

Run it only on instagram.com with local-storage permission, load it unpacked in developer mode, and keep the Node backend on 127.0.0.1:8787 with both keys configured. It never clicks Instagram controls.

What does the overlay actually show me?

Four signals behind View signals: harm, promotion, useful and sensitive percentages, each with the provider that led. The border is the summary, in four colours: green for potentially useful text, blue for low textual risk, orange for an Instagram ad label or a high Jev promotion probability, and red for content judged genuinely harmful. New feed articles and standalone Reel videos are covered automatically.

How does it avoid blocking ordinary sad content?

Non-graphic grief, illness, hospital, mortality and recovery themes are evaluated separately from graphic content or content encouraging harm, specifically to cut false red blocks. The backend test suite covers that false-positive case alongside Jev request construction, Gemini structured visual results, platform ad overrides, combined decisions and fallbacks, and runs with npm test from the backend folder.

Can I reveal content while still testing it?

Yes, and you should before trusting it. Reveal for this test exposes a single post, and Pause exposes the full feed temporarily, so you can inspect the prototype against real content without leaving it armed. Posts start covered by a safety placeholder rather than loading as usual.

CommunityJev Tools & CLIs

Instagram Safe Feed

Chrome proof of concept that sends Instagram post text to Jev and sampled images to Gemini, then overlays a colored border with expandable harm and promotion signals.

See highlights

Open github.com

Stars: 1
Language: JavaScript
Published by: Community
Author: Aagamsheth0601
Last verified: Oct 2026

Aagamsheth0601/instagram-safe-feed

User repository on GitHub

View on GitHub

Explainable Instagram feed and Reels classification using TypeSafe AI Jev and Gemini vision

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

This personal proof of concept initially covers Instagram feed posts and standalone Reels, sends visible text and metadata to TypeSafe's Jev model, and sends a downscaled post image or up to three sampled Reel frames to Gemini visual analysis. The percentages are model signals, not guarantees of safety. Instagram's own Ad, Sponsored, and Paid partnership labels take priority over AI classification.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Jev judges text and metadata while Gemini judges a downscaled image or up to three Reel frames.
Green, blue, orange, and red borders with expandable percentages and named providers.
Explicit Instagram ad labels override AI classification; high Jev promotion probability also maps to orange.
Local Node backend combines signals; failures and missing keys surface in the overlay with content covered.

Watch out

Personal research prototype, not a guaranteed safety product; visual analysis needs both API keys and the local backend running. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

using-typesafe-and-pinecone

Example collection pairing Pinecone retrieval with Jev judgments, including a reranking demo that narrows full-text candidates by natural-language criteria.