At the heart of conservation are the field staff who study and monitor ecosystems in challenging environments. Recent advances in AI models raise the question of whether LLM assistants could improve the experience of data collection for these staff. However, on-device AI deployment for conservation field work poses significant challenges, and is understudied. To address this gap, we conducted semi-structured interviews, surveys, and participant observation with partner conservancies in the Pacific Northwest and Namibia to better understand the field work context. We employ speculative methods through the lens of technology acceptance theory to critically analyze how on-device AI would affect field work, by developing an on-device transcription-language model pipeline, which we built atop of EarthRanger, a widely-used, open-source conservation platform. Our findings suggest that although on-device LLMs hold some promise for field work, the infrastructure required by current on-device models clashes with the reality of resource-limited conservation settings.

Many sensing tasks, such as plant stress sensing and blood pressure estimation, require co-located multi-modal measurements from two to five sensors at one site. RF backscatter enables low-power sensing, but existing tags usually support only one sensor; using multiple tags increases footprint and antenna coupling. We present Matrix, a fully-analog single-chain backscatter tag that supports multiple onboard sensors by multiplexing them into a composite voltage for transmission through one analog modulation chain. Unlike time-division polling, which introduces inter-sensor sampling offsets, or frequency-division, which requires separate chains, Matrix uses voltage-division multiplexing. Each sensor is encoded as a PWM waveform whose duty cycle represents the measurement, while amplitude enables multiplexing. Binary-weighted voltage-division weights make each active-sensor set uniquely invertible for reliable demultiplexing. The composite voltage is then converted into backscatter frequency shifts through the same chain. At the receiver, Matrix uses a Hidden Markov Model to recover per-sensor readings. Its ASIC consumes 25.56μW. A five-sensor prototype achieves 20 dB average reconstruction SNR at 30 kHz sampling, and we validate Matrix in plant sensing, health monitoring, and microphone-based direction finding.

Several methods have been developed to computationally predict cell-types for single cell RNA sequencing (scRNAseq) data. As methods are developed, a common problem for investigators has been identifying the best method they should apply to their specific use-case. To address this challenge, we present CHAI (consensus Clustering tHrough similArIty matrix integratIon for single cell-type identification), a wisdom of crowds approach for scRNAseq clustering. CHAI presents two competing methods which aggregate the clustering results from seven state-of-the-art clustering methods: CHAI-AvgSim and CHAI-SNF. CHAI-AvgSim and CHAI-SNF demonstrate superior performance across several benchmarking datasets. Furthermore, both CHAI methods outperform the most recent consensus clustering method, SAME-clustering. We demonstrate CHAI's practical use case by identifying a leader tumor cell cluster enriched with CDH3. CHAI provides a platform for multiomic integration, and we demonstrate CHAI-SNF to have improved performance when including spatial transcriptomics data. CHAI overcomes previous limitations by incorporating the most recent and top performing scRNAseq clustering algorithms into the aggregation framework. It is also an intuitive and easily customizable R package where users may add their own clustering methods to the pipeline, or down-select just the ones they want to use for the clustering aggregation. This ensures that as more advanced clustering algorithms are developed, CHAI will remain useful to the community as a generalized framework. CHAI is available as an open source R package on GitHub: https://github.com/lodimk2/chai.