Systems & research engineer working on reliable AI, power electronics, autonomous systems, and security-critical infrastructure.
I like problems where the abstraction can lie: model evaluations with leakage, uncertainty estimates that are not calibrated, network diagnostics that need real hardware, and protocols whose correctness has to survive adversarial conditions.
I am connecting layers that are usually discussed separately:
| Layer | Work | Question |
|---|---|---|
| Legacy energy | Darkhan modernization: a new turbine inside a plant commissioned in 1965 | Which inherited interfaces limit the upgrade? |
| Power conversion | VoltForge — physics-based GaN/SiC converter design with calibration gates | Does a better switch improve the complete power tree? |
| Specialized compute | LCA-1 (private, pre-FPGA) — a pinned bridge workload, golden model, and synthesizable dual-modulus arithmetic slice | Does measured end-to-end value justify a full engine? |
| Physical evidence | roce-preflight and hardware experiments | Which claims survive independent device state? |
The common rule: a component-level win is not a system-level win until interfaces, power, thermal behavior, and independent measurements agree.
- A new turbine inside a 1965 power plant — a 35 MW fifth unit, 13% instantaneous station use, and the interface tax of expanding legacy infrastructure.
- 197 passing tests, four real-hardware bugs — why synthetic confidence has to answer to independent device state.
- A benchmark can measure its own metadata — a public correction after a prompt audit found severe leakage in 13/24 benchmark examples.
- Upstream is a different kind of test — what merged third-party patches changed about the unit of correctness.
- VoltForge — vendor-neutral GaN/SiC power-converter design exploration; public accuracy claims are gated on measured-reference calibration.
- LCA-1 (private, pre-FPGA) — pinned ETP workload, golden model, and synthesizable ML-KEM/ML-DSA arithmetic slice; physical speedup and energy remain unproven.
- roce-preflight — RDMA/RoCE diagnostics with real-hardware validation; hardware CI found defects that unit tests missed.
- BRIDGE-bench — empirical AI-security evaluation work, including contamination controls and evaluator-validity analysis.
- active-materials-discovery — materials ML where the mean prediction was useful and the tested MC-dropout uncertainty signal failed calibration.
- Aiur — current autonomy/hardware experiment; design, simulation, controller, and acceptance-gate work, with physical recovery performance still explicitly unproven.
ML / scientific computing
- materialyzeai/matgl #801 — MC-dropout uncertainty estimation for pretrained materials models.
- materialyzeai/matgl #809 — autograd-correct, numerically safe
SoftExponentialactivation. - ml-explore/mlx-lm #1372 — corrected pathological XTC sampling defaults across library, CLI, and server.
Rust / systems
- alloy-rs/core #1105 — EIP-712 self-referential struct canonicalization in
dyn-abi. - uutils/coreutils #12327 — GNU-compatible
datetimezone re-zoning. - uutils/parse_datetime #284, #285, #287 — date parsing, negative Unix timestamp semantics, and timezone compatibility.
I build Suwappu, where I work on distributed systems, cryptographic protocols, transaction infrastructure, and adversarial reliability. I also maintain experiments in ML inference, scientific computing, networking, power conversion, and autonomous hardware.
Research & work · writing · research · X · layerinfinite@gmail.com