Projects
Selected machine learning systems, open-source tools, and datasets I have designed, trained, and shipped.
Bio-inspired cognitive architecture for LLM agents and robots.
Maxim provides homeostatic drives, persistent cross-session memory, reward learning, and retrieval-augmented generation (RAG) for LLM agents, with a simulation environment and robotics integration. Its components are computational models inspired by mammalian neuroscience, including the hippocampus, entorhinal cortex, nucleus accumbens, and attention networks. Maxim runs on the Reachy Mini robot with live visual input, motor commands, and human interaction.
Stack: Python (PyPI package pymaxim, Apache-2.0) · LLM agents · RAG · reward learning · Reachy Mini robot · Minecraft world bridge · ed25519-signed substrate sharing
Status/Result: Current release 1.3.1 on PyPI, documented at pymaxim.bio, with every claim pre-registered and tracked in a public evidence ledger. On a live Minecraft server with no LLM in the action path, a taught want transferred between independent agents (1.2) and a fear learned from in-game pain transferred to an agent that never felt it (1.3). On Reachy Mini hardware, the reward substrate learned to turn toward speech within ~10 trials (write-up).
Overview · Docs · Code · PyPI
Multimodal deep learning for measuring interpersonal synchrony.
synchronAI fuses video, audio, and fNIRS (functional near-infrared spectroscopy) time-series to detect when two people’s behavioral and physiological signals align during interaction. It gives psychology and neuroscience researchers an automated, quantitative measure of a phenomenon that has historically been difficult to assess objectively.
Stack: Multimodal deep learning over video, audio, and fNIRS time-series
Status/Result: In active development at the Laboratory for Child Brain Development, Washington University in St. Louis; current model reaches 0.72 AUC.
Code
Open-source Python package for hemodynamic response modeling in fNIRS.
HRfunc uses linear algebra and deconvolution to estimate neural activity and hemodynamic response functions (HRFs) directly from fNIRS data, rather than assuming a single canonical response. A hybrid tree–hash-table data structure backs a live peer-to-peer database that lets researchers share and reuse observed HRFs across studies and populations.
Stack: Python · pip-installable package (PyPI) · deconvolution · peer-to-peer HRF-sharing database
Status/Result: Published in Neurophotonics (2025); presented in an invited NIRx vendor webinar; maintained across 22 public releases with ~14K PyPI downloads.
Site · Code · PyPI · Paper
An open snowpack imagery dataset, a pretrained GAN, and a downstream avalanche-risk model.
The Rocky Mountain Snowpack dataset is a curated set of 4,040 snowpack images, each collected to preserve its location in time and space, released so the wider machine learning community can apply its methods to snow science. snowGAN was generatively pre-trained on this high-resolution imagery and serves as a transfer-learning backbone for AvAI, a multimodal avalanche-risk classifier. AvAI is delivered through a cross-platform mobile app for real-time inference and labeled field-data collection, and AvServ, a federated server that runs a check-out safety program connecting app users with search and rescue teams when they do not check in on time.
Stack: PyTorch / TensorFlow · GANs (progressive growing) · transfer learning · Hugging Face · cross-platform mobile app
Status/Result: Dataset published on Hugging Face (DOI 10.57967/hf/6143; peak of ~2,600 monthly downloads). Pretrained snowGAN models released as RMDig/snowGAN-core and RMDig/snowGAN-magnified-profile. AvAI is in development: model training and the mobile app are in progress.
Code · Dataset · DOI · snowGAN-core · snowGAN-magnified-profile · AvAI
Deep learning framework for fMRI (BOLD) data.
BOLDnet is a TensorFlow framework for rapid model prototyping on standardized fMRI datasets, supporting classification and regression on blood-oxygenation-level-dependent (BOLD) data. It includes custom spatial-attention layers and 4D volumetric data loaders.
Stack: TensorFlow · custom spatial-attention layers · 4D volumetric data loaders
Status/Result: Research prototype; originally developed as KleinNet in the Banich Lab at the University of Colorado Boulder.
Code
