Performance is a function of what you control. So I control all of it.

Research Software Engineer / Systems Architect. Bare-metal systems work — Rust, C, C++, CUDA where it matters; Python where it pays. I build my own stack, on my own hardware, with benchmarks I can defend.

The work spans four frontiers, with one thread running through them: physics computes by minimizing energy. I build the systems that exploit it.

• AI — agentic architectures, low-latency inference, distributed training
• Quantum — tensor networks, ground-state methods, quantum-inspired optimization
• Computational neuroscience — closed-loop BCI, attractor dynamics, energy-minimization models
• Energy / solar — sovereign DC-microgrid control, post-quantum trust anchors

The brain, the crystal, the optimizer, the controller — same physics, all relaxing to a minimum. The hardware is the only honest place to do that.

These four are the spine of the ecosystem — each one a domain bet, all bare-metal:

24/7 predictive DC-microgrid controller. Rust MPPT loop, CNN-LSTM forecasting, post-quantum trust anchors. Where the lights actually have to stay on.

Closed-loop BCI kernel. Sub-100µs CUDA DSP latency, compile-time-enforced neuroethics, post-quantum cryptography. Where milliseconds carry weight for the right reasons.

Multi-scale physics & quantum-biology simulation. Hand-written CUDA (sm_120) with an honest kernel-only roofline (3×→154× vs JAX-CPU) and tensor-network solvers for cryptochrome, EM fields, and multi-body dynamics.

Tensor-Train / MPS compression for high-order scientific and quantum-state data. GPU SVD, MPS↔circuit bridge. Cross-cuts the ecosystem; also the χ thermometer in DRIFT.

DRIFT reads optimization, self-assembly, self-replication, and neural memory (Hopfield, Nobel Physics 2024) as ground states of one Ising Hamiltonian — the unification thesis, made measurable:

How far is real hardware from the physical floor? Six orders of magnitude above the Landauer wall. That gap is the room left for computronium:

• DRIFT — A microscope for physical computation. Four faces, one Ising engine; tensor-network χ as the compute-density thermometer. P0–P7 done.
• TESSERA — Neural-guided real quantum annealing via tensor networks (MPS/TEBD); a GNN learns the schedule on a local GPU.
• BLACKWALL — Honest precision-spectrum GEMM roofline on Blackwell (sm_120): FP32 → FP4 measured, FP4 at 20× FP32 via cuBLASLt, anchored to the computed peak.
• EIGEN — Quantum-inspired threat modeling via transverse-field Ising models and a pure-Rust Lanczos solver.
• KINECT-NIR — Real-time IR object detection & tracking. CUDA CA-CFAR kernel + TensorRT INT8 + Kalman filtering.

Several larger systems are in private development and can be discussed in conversation.

Systems — Rust · C · C++ · CUDA (sm_120 Blackwell) · CUDA-Q · Assembly Quantum — Cirq · tensor networks (TT / MPS / TEBD) · QUBO / Ising · Lanczos / Krylov AI / ML — Python · PyTorch · cuBLAS · cuBLASLt · TensorRT · Flash Attention HPC / Infra — SLURM · NCCL · Arch Linux · bare-metal Sovereignty — Post-quantum primitives · compile-time safety · auditable code

Bare-metal · local-first · honest benchmarks · no cloud by default.