# AI use cases for Unbiased Strata (robertkarl/UnbiasedStrata) Repo: [robertkarl/UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata) | Use cases: 4 | Page: https://dedupedai.com/use-cases/robertkarl-unbiasedstrata/ UnbiasedStrata is a C++ fork of PolyStrata that builds two local inference engines and a front end, unbiased-strata, which serves an OpenAI-style chat API for one NVIDIA GPU on Linux. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Serve a local chat API on a machine with no network: https://dedupedai.com/use-cases/robertkarl-unbiasedstrata/serve-local-chat-api-offline/ - Ask a local model questions about screenshots: https://dedupedai.com/use-cases/robertkarl-unbiasedstrata/ask-questions-about-screenshots-locally/ - Test the server front end without a GPU or model: https://dedupedai.com/use-cases/robertkarl-unbiasedstrata/check-front-end-without-gpu/ - Prepare a Flash-Next pack for the strata engine: https://dedupedai.com/use-cases/robertkarl-unbiasedstrata/prepare-flash-next-pack-offline/ ## How to serve a local chat API on a machine with no network with Unbiased Strata Repo: [robertkarl/UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata) | Difficulty: Advanced | Time: 3 hours | Basis: From the README Build the engines and run an OpenAI-style chat endpoint on one NVIDIA GPU with no downloads at build or start. **Who it is for:** A developer or ops engineer with an NVIDIA Linux box and a GGUF model who needs a private chat API that works offline. **The problem:** Installers and runtime downloads make local model servers hard to trust and hard to run on isolated machines. **The outcome:** A running unbiased-strata process on 127.0.0.1 that answers /v1/chat/completions from the local model. ### Steps 1. Run `git clone https://github.com/robertkarl/UnbiasedStrata` then `cd UnbiasedStrata`. (inferred, check the README) 2. Run `git submodule update --init` to fetch the pinned llama.cpp fork. 3. Run `make build` to produce build/bin/strata-poly, build/bin/strata, libstrata_kq.so and unbiased-strata. 4. Run `make test` to check the front end against a fake engine, with no GPU or model needed. 5. Start the server with `build/bin/unbiased-strata --engine build/bin/strata-poly -m /models/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf --host 127.0.0.1 --port 18091 --alias qwen3.6-35b-a3b-ud-q3_k_m -- --model /models/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf --max-context 32768 --expert-profile /models/expert-profile-q3km.txt --tail-skip 6`. (inferred, check the README) ### Prompt to paste into your agent ```text In the UnbiasedStrata repo at , run make build and then make test, and report whether the test suite passed. Then start unbiased-strata with the model at on port 18091 and confirm that GET /health returns a success response. ``` ### Caveat The README's model and expert-profile paths are examples, so you must supply your own GGUF files, and the target is one NVIDIA GPU on Linux x86-64 only. 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/robertkarl-unbiasedstrata/serve-local-chat-api-offline/ ## How to ask a local model questions about screenshots with Unbiased Strata Repo: [robertkarl/UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata) | Difficulty: Advanced | Time: 2 hours | Basis: From the README Attach a projector file so the local server can read pictures such as dashboards, without sending them to a cloud API. **Who it is for:** An analyst or engineer who has dashboard screenshots they cannot upload to a hosted service. **The problem:** Reading images with a hosted model means sending the pictures off the machine. **The outcome:** The local server accepts picture input and answers questions about the image, with encoding done on the same machine. ### Steps 1. Build and start the server as in the chat API setup, then add `--mmproj /models/mmproj-F16.gguf` and `--threads N` before the `--` separator. 2. Keep the default picture size of 300 rows, or change it with `--image-max-tokens`. 3. Send a chat request to http://127.0.0.1:18091/v1/chat/completions with the screenshot attached as an image part, using the model alias you set with --alias. (inferred, check the README) 4. Check the answer against the screenshot yourself before relying on it. ### Prompt to paste into your agent ```text In the UnbiasedStrata repo at , start unbiased-strata with the projector at and report the server log line that confirms the projector loaded. Then send the image at to /v1/chat/completions and report the reply. ``` ### Caveat The README reports one 1920x1080 dashboard read in about 15 seconds of encoding on a Ryzen 7 1700, so timing will differ on other hardware, and the projector must match the model. 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/robertkarl-unbiasedstrata/ask-questions-about-screenshots-locally/ ## How to test the server front end without a GPU or model with Unbiased Strata Repo: [robertkarl/UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README Run the front end test suite against a fake engine to confirm a build works before any model is loaded. **Who it is for:** A developer who changes the C++ front end or builds on a machine with no GPU and no model files. **The problem:** Full inference tests need a GPU and large model files, which slows down checking a build. **The outcome:** A pass or fail result for the front end tests that does not need hardware or model files. ### Steps 1. Run `git submodule update --init` in the cloned repo. 2. Run `make build`. 3. Run `make test` and read the result. ### Prompt to paste into your agent ```text In the UnbiasedStrata repo at , run make build and make test and report the final pass or fail count. Confirm the tests ran against the fake engine and not a real model. ``` ### Caveat make build still needs the CUDA toolchain on Linux x86-64, so this is not a no-hardware build of the engines themselves. 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/robertkarl-unbiasedstrata/check-front-end-without-gpu/ ## How to prepare a Flash-Next pack for the strata engine with Unbiased Strata Repo: [robertkarl/UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata) | Difficulty: Advanced | Time: 4 hours | Basis: From the README Build the pack directory and MTP draft directory once so the strata engine can serve Qwen3.8-Flash-Next. **Who it is for:** An operator who wants the Flash-Next IQ3_XXS model served by strata and has the upstream GGUF shards on hand. **The problem:** The strata engine needs a pack directory and a draft directory that are not made by the front end, and the steps are easy to get wrong. **The outcome:** A pack directory and MTP draft directory ready to pass to strata through the --pack and --mtp options. ### Steps 1. Install the Python tool requirements listed in tools/requirements.txt (numpy and regex), with gguf-py from the submodule. 2. Run `python tools/iq_pack.py --gguf --out` with your GGUF path and output directory (add --compat-bf16 for Unsloth files). 3. Copy the published MTP directory from https://huggingface.co/robertkarl/UnbiasedStrataPacks instead of fetching and quantizing it, or run `STRATA_SOURCE=huggingface python tools/mtp_fetch.py fetch --out`. 4. Run `python tools/mtp_pack.py --src --experts q2_0 --out /mtp-q2_0.gguf` and `python tools/mtp_rt.py --gguf /mtp-q2_0.gguf --out /rt`, then `cp data/draft_vocab.bin /rt/draft_vocab.bin`. ### Prompt to paste into your agent ```text In the UnbiasedStrata repo at , prepare the Flash-Next pack from the GGUF shard at and report the output directory and its file list. Verify that draft_vocab.bin is present in the MTP directory before reporting success. ``` ### Caveat The README's commands leave some arguments blank, so fill in the paths yourself, and the published MTP files are the ones the author hashed, so check your copy against the listed hashes. 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/robertkarl-unbiasedstrata/prepare-flash-next-pack-offline/