Raven Nicole Roberts

I built a mineral identification system designed for serious collectors, geology students, and field researchers—not a generic image classifier with marketing veneer. The architecture fuses multiple modalities: visual features (color, luster, crystal habit) extracted from user photos; optional UV fluorescence imaging; locality priors derived from USGS MRDS and mindat.org data indexed by H3 spatial cells; and confidence calibration that explains what drove the prediction and what reduced confidence. The solution is paired with a physical specimen line where each shipped specimen carries a QR code linking to its ground-truth entry, generating consented in-the-wild validation data. It was built mobile-first with a dark, scientific aesthetic distinct from generic SaaS design.

This is a reinforcement learning agent framework built on Clifford geometric algebra, where rotational and Lorentz equivariance are native to the representation rather than enforced as a loss-function constraint. I implemented Cl(3,0) and Cl(3,1) layers from scratch in PyTorch, with the geometric product as the core nonlinearity, wrapped in a Gymnasium-compatible agent interface. The first benchmark task is plasma instability control via BOUT++ as the simulation back end—chosen because plasma dynamics naturally live in the bivector field structure that Clifford algebra exposes. It achieves equivariance at roughly the same parameter count as a vanilla MLP, which matters for deployment on real hardware. The solution includes production-quality engineering: typed APIs, tested algebra primitives, and clean separation between mathematical primitives and agent logic.

A variational quantum eigensolver (VQE) approach to classifying LIGO gravitational wave signals from real detector strain data, executed on IBM Quantum hardware via PennyLane. The architecture uses parameterized quantum circuits as the feature map, trained with classical optimizers, and benchmarked against classical baselines on detection sensitivity and signal-to-noise robustness. The work explores whether quantum machine learning offers a genuine advantage on physics-native problems where the data structure (non-stationary noise, weak signals, high-dimensional time-frequency representations) might align with quantum representational capacity. I co-authored research with a UCSB PhD candidate, with an arXiv preprint in preparation.

I developed a multimodal AI knowledge system that ingests text, images, video, audio, medical scans, mind maps, and 3D data, forming connections between related items and surfacing latent patterns. I built the solution on a hybrid architecture where every connection has two parallel representations: a symbolic edge in a property graph for fast, legible queries, and a multivector in a Clifford geometric algebra for compositional and analogical reasoning that the symbolic layer can't express. The solution uses modality-specific encoders (DINOv2, RadDINO, CLAP, BGE, Uni3D) rather than a lossy unified CLIP-style embedding. Time is a first-class axis alongside semantic similarity, so queries like what was I working on the day I got that scan? resolve naturally.

The interface is a TouchDesigner-inspired node-graph canvas where proposed connections are visible, editable, and rejectable, a choice made because chat-only interfaces hide hallucinated connections in personal knowledge systems. RSS and read-later integration (Miniflux and Wallabag) brings external content into the same graph. The solution is self-hostable with an optional local LLM endpoint via Ollama for fully air-gapped operation.