cruiseresearchgroup/ZARA — reverse-engineered prompt

Build me a working research prototype of ZARA, a training free system that can recognize human activities from wearable motion sensor time series by using an LLM agent instead of training a classifier. I want it to feel transparent, not like a black box, so the result for each prediction should include a short human readable rationale based on retrieved signal evidence and simple feature based prior knowledge.

Please set it up around the two example flows in the repo, one for UCI HAR and one for the Shoaib dataset. Each dataset should have a clear notebook sequence for preprocessing, generating feature importance based knowledge, and then running Gemini inference. For the Shoaib example, include the placement aware retrieval and rank fusion behavior. For the UCI example, keep the simpler single retrieval index setup.

Make it easy to run with an API key in an environment variable, and add just enough documentation so someone can open the notebooks and reproduce the end to end pipeline. If anything is unclear, look up the paper or current docs online.