Andrew Jewell Sr.

We present TMC (Thermal Momentum Control), a nonlinear feedback controller for heterogeneous HVAC equipment. TMC combines a signed-demand PI core with wrong-direction-gated velocity feedforward, a thermal-coasting brake zone derived from first-order plant coasting distance, back-calculation anti-windup augmented by overshoot-triggered integrator reset, online adaptation of the thermal time constant via a rate-based midpoint-velocity estimator, and runtime Lyapunov monitoring. The controller is implemented as a single Rust core shared unchanged across ten HVAC equipment classes. We provide a closed-loop stability analysis via second-order ODE reduction, and validate against a Ziegler-Nichols tuned PID baseline. On a cascaded two-stage thermal plant, TMC limits overshoot to 0.34°F where an aggressively-tuned PID overshoots by 2.01°F—a 6× reduction. We further present field-validation results from a production deployment of 113 active HVAC units across six buildings via the NexusEdge control platform.

We present Trident, an end-to-end implementation of a BitNet b1.58 ternary language model in pure Rust, including the ternary linear layer, absmean weight quantization, Straight-Through Estimator backward pass, training loop, checkpoint serializer, and generation utility. The implementation contains no Python in the dependency tree and no C++ wrapper around libtorch; all components are built on the AxonML framework with native CUDA acceleration. We train a 616{,}448-parameter model on the complete works of Shakespeare (~5.4M characters) for 100 steps, observing cross-entropy loss decrease from 7.9 to 2.61 (perplexity 13.6) with ternary sparsity stabilizing near 25%. We report storage compression ratios of 9.71× to 11.99× for the Shakespeare configurations and 2.89× for a 162M-parameter variant. Released under MIT and Apache-2.0.

We present AxonML (v0.6.1), an open-source deep learning framework implemented entirely in Rust, spanning 22 crates and 226{,}373 lines of code. The framework provides a complete ML stack: N-dimensional tensor operations with CUDA GPU acceleration via 15 PTX kernel modules, reverse-mode automatic differentiation with gradient checkpointing and automatic mixed precision, 41 neural network layer types including state-of-the-art mechanisms (rotary embeddings, grouped-query attention, mixture-of-experts, differential attention, ternary quantized layers), 5 optimizers, 7 learning rate schedulers, and 7 loss functions. AxonML includes nine LLM architectures, six quantization formats with 1.58-bit ternary support, distributed training (DDP, FSDP, pipeline parallelism), ONNX import/export, and a 33-subcommand CLI. The workspace contains 2{,}285 tests. Dual-licensed under MIT and Apache-2.0.