Jev directory
CommunityJev Demos & Experiments
MedGemma Ophthalmology + Jev Demo
A research demo that runs a fine-tuned MedGemma 1.5 4B adapter on retinal fundus images in Colab, passes the generated findings through Jev as a text quality gate, and drafts ophthalmologist and patient summaries via OpenRouter.
- Stars
- 0
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- Fine-tuned MedGemma 1.5 4B LoRA adapter with 272 adapter tensors across 136 language-only projections
- Jev runs through OpenRouter and applies a text-based quality gate that rejects readability below 0.70
- Runs on a Colab T4 GPU behind a Gradio interface; the first run downloads the gated base-model weights
- Produces separate ophthalmologist and plain-language patient summary drafts using OpenAI models via OpenRouter
- The author states it is not a clinical decision system and results require qualified clinician review
Questions about this entry
Short answers drawn from the linked project or page.
Open the Colab notebook, choose Runtime > Change runtime type > T4 GPU, install the listed packages, upload model/colab_demo.py and your .env, then run the script and open the printed Gradio share link. The first run downloads the gated MedGemma base-model weights.
You need an OPENROUTER_API_KEY for Jev and the generation models, plus an HF_TOKEN from a Hugging Face account that accepted the gated MedGemma terms and can access the trained adapter. No separate TypeSafe or OpenAI key is needed, but API credits are required.
Uploaded retinal images stay on the Colab GPU where MedGemma runs; the generated findings text is sent to Jev and to OpenAI models through OpenRouter. Only de-identified images should be used, and the uploaded .env should be deleted after the demonstration.
Jev returns REJECTED_QUALITY_GATE when text usability falls below 0.70; no grade or grade-specific follow-up is accepted and deterministic review notices replace the GPT drafts. INVALID_MODEL_OUTPUT or JEV_ERROR also stops any diagnostic summary.
No. Jev evaluates the text findings generated by the MedGemma adapter, not the image independently. Missing quality commentary can fail the gate without proving the photograph itself is unusable.
The author states it is a research demonstration, not a clinical decision system or confirmed diagnosis. Model confidence and urgency are not validated clinical metrics, and no clinician-approved referral protocol is configured.
The published demo is a Colab notebook using a T4 GPU; hosting adapter weights on Hugging Face is not remote inference, and the model runs on the Colab GPU rather than the user's local computer. The training export is an IPython notebook intended for Colab.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Privacy-first Next.js app that sorts saved X bookmarks into Top picks, Needs human look, and Skip using one Jev call per bookmark. Composite scoring, confidence-gated routing, and a deterministic mock mode that runs without an API key.
Jev Demos & Experiments#demo#typescript#routing
CommunityPirozshki/jev-bookmark-triage
A personal Docker app that collects stock news from 29 RSS outlets and Finnhub every 30 minutes, then uses Jev to score each article for relevance, impact, and whether it reads as bullish or bearish, showing it on a web timeline.
Jev Demos & Experiments#madewithjev#github#classification
Communityanjfgkfkrh/Ripple
A Beam-hosted playground for comparing open-source alternatives to Jev and running any of them as an API. You edit a state and typed questions in the browser, then call the same model through an OpenAI-compatible endpoint with a workspace token.
Jev Demos & ExperimentsPlayground#community#catalog-import#madewithjev
Community
A playable Three.js driving simulator where Jev-powered autopilot picks steering and speed up to four times per second: compact tables of eligible paths, traffic and road boundaries go to Jev while geometry and control math stay local.
Jev Demos & Experiments#community#catalog-import#cobanov
Community
standardagents/jevpilotJavaScript
standardagents/jevpilotJavaScriptLocal macOS computer-use automation on Apple silicon using on-device laya typed decisions and CUA-S1-FORMS for filling, via the Accessibility API.
Jev Demos & Experiments#community#browser#local
Community
FluidInference/FluidUseSwift
FluidInference/FluidUseSwiftAn agent skill that reproduces Jev-style single-step logits over indexed browser controls using local Qwen weights for offline browser use.
Jev Demos & Experiments#community#browser#open-models
Community
APUS-AI-Lab/fast-browser-usePython
APUS-AI-Lab/fast-browser-usePythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.