# AI use cases for RSI Jev (Shanghua-Gao/RSI-Jev) Repo: [Shanghua-Gao/RSI-Jev](https://github.com/Shanghua-Gao/RSI-Jev) | Use cases: 5 | Page: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/ RSI-Jev is a set of typed-decision models (yes/no, pick-one and rubric questions) that AI agents trained in a self-improving loop, with the checkpoints, code and failed experiments published. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Serve a decision model behind a local API: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/serve-decision-model-local-api/ - Make a typed decision in Python without a server: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/decide-in-python-without-server/ - Run the 4B decision model on a Mac: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/run-4b-model-on-apple-silicon/ - Point an existing Jev client at this server: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/point-existing-jev-client-at-server/ - Check failed experiments before planning a training run: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/check-failed-experiments-before-training/ ## How to serve a decision model behind a local API with RSI Jev Repo: [Shanghua-Gao/RSI-Jev](https://github.com/Shanghua-Gao/RSI-Jev) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Run v6.1-VL 4B locally and get calibrated yes/no or pick-one answers over HTTP. **Who it is for:** A developer whose app needs a fast yes/no or pick-one answer on text or images, with no per-call generation cost. **The problem:** Calling a general chat model for every routing or classification decision is slow and spends tokens on reasoning the task does not need. **The outcome:** A local server that returns a probability for each option, plus the layer that answered and its confidence. ### Steps 1. Install the package with `pip install "rsi-jev[fast,vision] @ git+https://github.com/Shanghua-Gao/RSI-Jev"` 2. Start the server with `rsi-jev serve v6.1-vl-4b --effort auto --port 8000` 3. Send a test question with `curl localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{"model": "jev-latest", "state": "I was charged twice. Please refund.", "questions": {"refund": {"type": "noul", "instructions": "Does the user request a refund?"}}}'` (inferred, check the README) 4. Check that the response reports the layer that answered and its confidence ### Prompt to paste into your agent ```text In the Shanghua-Gao/RSI-Jev repo, install the package with pip, then run `rsi-jev serve v6.1-vl-4b --effort auto --port 8000`. Send the about the text to localhost:8000/v1/systemone and return the probability for each option. Verify that the response includes the layer that answered and a confidence value. ``` ### Caveat The public Decision Index scores are for full runs of the default configuration, so your own results on different questions will vary. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/serve-decision-model-local-api/ ## How to make a typed decision in Python without a server with RSI Jev Repo: [Shanghua-Gao/RSI-Jev](https://github.com/Shanghua-Gao/RSI-Jev) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Call the Decider class in a script to score options for text or images. **Who it is for:** A Python developer building a batch job or pipeline that needs decisions on documents, chats or images. **The problem:** Standing up a separate server just to classify a set of records adds setup work for a one-off or scripted task. **The outcome:** A script that returns a decision for each named question, with images passed in when needed. ### Steps 1. Install the package with `pip install "rsi-jev[fast,vision] @ git+https://github.com/Shanghua-Gao/RSI-Jev"` 2. Create a Python file that imports Decider with `from rsijev import Decider` 3. Load the model with `d = Decider("v6.1-vl-4b")` 4. Call `d.decide(...)` with the state text, a dictionary of typed questions and an images list, then run the script (inferred: with `python .py`) (inferred, check the README) ### Prompt to paste into your agent ```text Using the Shanghua-Gao/RSI-Jev package, write a Python script that loads Decider with v6.1-vl-4b and calls decide() on each record in with a noul question about . Output one line per record with the option probabilities. Verify that every record gets a result and that image inputs are passed only when present. ``` ### Caveat The README example passes one image path, so check the Decider interface in the repo before running it on a large batch. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/decide-in-python-without-server/ ## How to run the 4B decision model on a Mac with RSI Jev Repo: [Shanghua-Gao/RSI-Jev](https://github.com/Shanghua-Gao/RSI-Jev) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Serve the MLX build of v6.1-VL 4B on Apple silicon using the MLX backend. **Who it is for:** A developer or analyst on an Apple silicon Mac who wants to test decision models without a dedicated GPU. **The problem:** Most model serving setups assume an NVIDIA GPU, which a Mac does not have. **The outcome:** A local server on a Mac that answers decision questions with the 8-bit or 4-bit MLX build. ### Steps 1. Install the MLX extras with `pip install "rsi-jev[mlx,vision] @ git+https://github.com/Shanghua-Gao/RSI-Jev"` 2. Start the 8-bit build with `rsi-jev serve v6.1-vl-4b-mlx-8bit --backend mlx` 3. Read docs/mac.md in the repo for the 4-bit build, which is smaller, and for setup details 4. Send a test request to the local server as shown in the repo's quickstart ### Prompt to paste into your agent ```text In the Shanghua-Gao/RSI-Jev repo, install the mlx extras with pip, then run `rsi-jev serve v6.1-vl-4b-mlx-8bit --backend mlx`. Follow docs/mac.md and send one decision request about to the local server. Return the answer and the layer reported. Verify that the server runs on Apple silicon and returns a response without errors. ``` ### Caveat The README describes the MLX builds only for the 4B model, so do not expect the 27B model to run on a Mac from these instructions. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/run-4b-model-on-apple-silicon/ ## How to point an existing Jev client at this server with RSI Jev Repo: [Shanghua-Gao/RSI-Jev](https://github.com/Shanghua-Gao/RSI-Jev) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Switch a Jev client's base URL to a local RSI Jev server without rewriting code. **Who it is for:** A team already using a Jev client that wants to run the models on its own hardware. **The problem:** Changing the client code for a new model host costs time and risks breaking working integrations. **The outcome:** The existing client gets decisions from the local server after only the base URL changes. ### Steps 1. Start the server with `rsi-jev serve v6.1-vl-4b --effort auto --port 8000` 2. Set the Jev client's base URL to http://localhost:8000 in its configuration (inferred, check the README) 3. Keep the model name as jev-latest, as shown in the quickstart request 4. Send one request through the client and check that the response includes the layer and confidence ### Prompt to paste into your agent ```text Given an existing Jev client in , point its base URL to localhost:8000 after running `rsi-jev serve v6.1-vl-4b --effort auto --port 8000` from the Shanghua-Gao/RSI-Jev repo. Run one existing client request and return the response. Verify that the client works with no other code changes. ``` ### Caveat The README says an existing Jev client works once its base URL points here, but it does not list which client features are covered, so test your own calls first. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/point-existing-jev-client-at-server/ ## How to check failed experiments before planning a training run with RSI Jev Repo: [Shanghua-Gao/RSI-Jev](https://github.com/Shanghua-Gao/RSI-Jev) | Difficulty: Beginner | Time: 15 minutes | Basis: From the README Read the published failed experiments to avoid repeating dead ends. **Who it is for:** A researcher or ML engineer planning a fine-tune or a training change for a decision model. **The problem:** Repeating a training idea that already failed wastes GPU time and delays a result. **The outcome:** A short list of past experiments on your topic, with the reason each one failed as the repo records it. ### Steps 1. Open EXPLORE.md in the repo to see all 496 experiments and why each failed 2. Open docs/history.md for the experiments that made a release 3. Open docs/rl.md for the reinforcement-learning arms 4. Open the release records in versions/ for the numbers behind a specific release ### Prompt to paste into your agent ```text In the Shanghua-Gao/RSI-Jev repo, read EXPLORE.md, docs/history.md and docs/rl.md and list every experiment that relates to . For each one, return the result and the failure reason as written in the repo. Verify that each item cites the file where it appears. ``` ### Caveat The 27B score of 65.64 is on the public set and its official score is still pending, so compare it with other results only with that caveat in mind. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/shanghua-gao-rsi-jev/check-failed-experiments-before-training/