Software engineer focused on distributed systems, cloud infrastructure, and AI.
Neural Network Observability (arXiv) Architecture and training decisions determine how observable an LLM is. Transformer activations carry decision-quality signals that output confidence misses; training can preserve or erase them during convergence, even as predictive performance improves. Across 26 models in 7 families, configuration and training recipe predict whether the signal survives. The signal lives in a low-variance geometric subspace that some configurations preserve and others lose. Signal engineering at training time is a design axis alongside loss and capability. Built with Python and PyTorch.
NxtChess (demo) Multiplayer chess platform with custom AI playstyles, anonymous play via shareable links, and OAuth 2.0 auth. Built with SolidJS, Go, PostgreSQL, Redis, and WebSockets.
Compass Autonomous agent architecture for real-time decision-making in a 3D MMORPG. Layered decision stack with GOAP planning, Monte Carlo robustness gating, Bayesian online learning, and JPS/A* pathfinding. Built with pure Python 3.14, zero external dependencies.
Material Type Prediction with WiFi/BLE Early wireless sensing experiment using commodity WiFi/BLE RSSI to detect objects and infer material class.
Carmichael, T. "Architecture Determines Observability in Transformers." arXiv preprint, 2026. [arXiv] [PDF] [Code] [Zenodo]
Dai, G., Paluri, P., Carmichael, T., Cheng, A., Miikkulainen, R. "Work-in-progress: Leveraging the Selfless Driving Model to Reduce Vehicular Network Congestion." IEEE Real-Time Systems Symposium (RTSS), 2019.
Email: hello@tdosmail.com