# AI use cases for LLM2 Jev (Yinsongxu/LLM2Jev) Repo: [Yinsongxu/LLM2Jev](https://github.com/Yinsongxu/LLM2Jev) | Use cases: 5 | Page: https://dedupedai.com/use-cases/yinsongxu-llm2jev/ LLM2Jev is an open-source Python project that turns a local language model or multimodal model into a decision model that scores options, rubrics, yes/no questions, and image boxes using prefill only, with no training. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Route support tickets to a department with probabilities: https://dedupedai.com/use-cases/yinsongxu-llm2jev/route-support-tickets-to-departments/ - Score drafts against a rubric and yes/no checks: https://dedupedai.com/use-cases/yinsongxu-llm2jev/score-drafts-against-a-rubric/ - Pick the next game action from an image: https://dedupedai.com/use-cases/yinsongxu-llm2jev/pick-next-action-from-a-screenshot/ - Locate a target object in an image with boxes: https://dedupedai.com/use-cases/yinsongxu-llm2jev/locate-targets-in-images-with-boxes/ - Run decisions offline on a Mac with MLX: https://dedupedai.com/use-cases/yinsongxu-llm2jev/run-decisions-offline-on-apple-silicon/ ## How to route support tickets to a department with probabilities with LLM2 Jev Repo: [Yinsongxu/LLM2Jev](https://github.com/Yinsongxu/LLM2Jev) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Score a customer message against department options and get a probability for each one. **Who it is for:** A support lead or an AI agent that triages incoming messages and needs a routing decision for each one. **The problem:** Routing messages by hand is slow, and a single hard label hides how sure the model is. **The outcome:** A department probability for each message, read from the Choice result, plus a chat endpoint on the same service for follow-up replies. ### Steps 1. `git clone https://github.com/Yinsongxu/LLM2Jev.git` 2. `cd LLM2Jev` 3. `uv sync --extra sglang` 4. `source .venv/bin/activate` 5. `llm2jev-serve --backend sglang --model-path /path/to/model --served-model-name local-model --host 127.0.0.1 --port 30000`, then run the client example from the README with your ticket text in `state` (inferred, check the README) ### Prompt to paste into your agent ```text Using the LLM2Jev repo, start the SGLang service with the model at , then call client.systemone.create with the ticket text and a Choice over the departments . Return the probability for each department as JSON. Verify that the probabilities sum to about 1 before reporting. ``` ### Caveat The README's quick start needs an NVIDIA GPU on Linux for SGLang, and the model must be one that SGLang supports. 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/yinsongxu-llm2jev/route-support-tickets-to-departments/ ## How to score drafts against a rubric and yes/no checks with LLM2 Jev Repo: [Yinsongxu/LLM2Jev](https://github.com/Yinsongxu/LLM2Jev) | Difficulty: Intermediate | Time: 1 hour 30 minutes | Basis: From the README Rate a text on rubric criteria and get yes/no probabilities for checks like policy compliance. **Who it is for:** A content editor or a QA agent that reviews drafts against a fixed checklist before publishing. **The problem:** Reviewers apply rubrics inconsistently, and generated text needs a repeatable first pass. **The outcome:** Rubric scores from Score and yes/no probabilities from Noul, returned as typed Python objects or JSON in one request. ### Steps 1. `git clone https://github.com/Yinsongxu/LLM2Jev.git` 2. `cd LLM2Jev` 3. `uv sync --extra sglang` 4. `llm2jev-serve --backend sglang --model-path /path/to/model --served-model-name local-model --host 127.0.0.1 --port 30000` (inferred, check the README) 5. Build one request that mixes a Score rubric and a Noul yes/no question, following the Usage guide section on the Python client (inferred, check the README) ### Prompt to paste into your agent ```text Using the LLM2Jev repo, call systemone.create on the running service with the draft text , a Score rubric for , and a Noul yes/no check for . Return the rubric scores and yes/no probabilities as JSON. Verify that each score is returned for every rubric item before reporting. ``` ### Caveat The README lists Score and Noul as question types but its quick start shows only Choice, so check the Usage guide for the exact Score and Noul arguments. 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/yinsongxu-llm2jev/score-drafts-against-a-rubric/ ## How to pick the next game action from an image with LLM2 Jev Repo: [Yinsongxu/LLM2Jev](https://github.com/Yinsongxu/LLM2Jev) | Difficulty: Advanced | Time: 2 hours | Basis: From the README Choose the best next move from a board image using a multimodal Choice question. **Who it is for:** A researcher or game developer testing whether a multimodal model can choose a good single action from a picture. **The problem:** Checking action choices from images usually needs custom glue code around a model. **The outcome:** Option probabilities for each candidate action, based on the image and a text state, from the single-step evaluator pattern. ### Steps 1. `git clone https://github.com/Yinsongxu/LLM2Jev.git` 2. `cd LLM2Jev` 3. `uv sync --extra sglang` 4. `llm2jev-serve --backend sglang --model-path /path/to/model --served-model-name local-model --host 127.0.0.1 --port 30000` (inferred, check the README) 5. Run `python examples/valen_sokoban_single_step.py` after the service is up, and check the Multimodal inputs guide for image arguments (inferred, check the README) ### Prompt to paste into your agent ```text Using the LLM2Jev repo, run the single-step Sokoban evaluator against the running SGLang service on with the multimodal Choice pipeline. Return the chosen action and option probabilities for each image as JSON. Verify that the output count matches the number of images before reporting. ``` ### Caveat The README says the evaluator uses the SGLang backend and reports accuracy on 500 Sokoban questions, but the exact run command and image input format are not in the README excerpt. 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/yinsongxu-llm2jev/pick-next-action-from-a-screenshot/ ## How to locate a target object in an image with boxes with LLM2 Jev Repo: [Yinsongxu/LLM2Jev](https://github.com/Yinsongxu/LLM2Jev) | Difficulty: Advanced | Time: 1 hour 30 minutes | Basis: From the README Get normalized bounding boxes for a named target in an image using the Box question type. **Who it is for:** A developer building an image review or screen-analysis tool who needs object locations from a model. **The problem:** Getting coordinates from a model usually means parsing free text, which is fragile. **The outcome:** Normalized bounding boxes returned as typed objects, which can be combined with Choice or Score in one request. ### Steps 1. `git clone https://github.com/Yinsongxu/LLM2Jev.git` 2. `cd LLM2Jev` 3. `uv sync --extra sglang` 4. `llm2jev-serve --backend sglang --model-path /path/to/model --served-model-name local-model --host 127.0.0.1 --port 30000` (inferred, check the README) 5. Follow the Box guide in docs/en/multimodal.md to send an image with a Box question for the target (inferred, check the README) ### Prompt to paste into your agent ```text Using the LLM2Jev repo, send the image and a Box question for to the running service. Return the normalized bounding box as JSON. Verify that coordinates fall between 0 and 1 before reporting. ``` ### Caveat The README says Box generates normalized coordinates, but it does not give the request format, so read the Box guide before writing code. 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/yinsongxu-llm2jev/locate-targets-in-images-with-boxes/ ## How to run decisions offline on a Mac with MLX with LLM2 Jev Repo: [Yinsongxu/LLM2Jev](https://github.com/Yinsongxu/LLM2Jev) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Score choices in Python on Apple Silicon without a server, using the MLX backend. **Who it is for:** A developer on an Apple Silicon Mac who wants local decisions without an NVIDIA GPU. **The problem:** Many local decision setups assume an NVIDIA GPU and a running server. **The outcome:** Choice or Score results from an offline Python call on the MLX backend. ### Steps 1. `git clone https://github.com/Yinsongxu/LLM2Jev.git` 2. `cd LLM2Jev` 3. Install the MLX dependencies using the Installation guide in docs/en/installation.md, then `uv sync` in the repo (inferred, check the README) 4. `source .venv/bin/activate` 5. Open the Offline: Python API section of docs/en/usage.md and run its example with your model path (inferred, check the README) ### Prompt to paste into your agent ```text Using the LLM2Jev repo on Apple Silicon, run the offline Python API with the MLX backend on the model at , asking a Choice question over for the input . Return the option probabilities as JSON. Verify that the probabilities sum to about 1 before reporting. ``` ### Caveat The README names MLX for Apple Silicon, but the exact MLX install steps and model requirements are only in the linked Installation and Usage guides. 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/yinsongxu-llm2jev/run-decisions-offline-on-apple-silicon/