Electrical & Embedded Systems Engineer · Modeling, Simulation and GPU Computing
From transformer windings to CUDA kernels — 16+ years of engineering that has to work on real hardware.
Electrical engineering fundamentals earned designing power equipment to IS/BS standards, then sixteen years of modeling, simulation and real-time systems — industrial machine vision, medical imaging, automotive ADAS — and a year authoring and reviewing GPU-programming evaluations for frontier AI models.
Most software engineers working on AI have never sized a radiator or calculated oil quantity for a 33 kV transformer. Most engineers who have, never made it to CUDA. The overlap is where I am useful.
I started as a design engineer for power equipment — power, earthing and distribution transformers and reactors up to 5 MVA at 33 kV, designed to customer specification and to IS and BS standards. Winding and core-cutting instructions, oil quantity and conservator selection, cooling and radiator calculations, and the AutoCAD drawings the shop floor actually built from. That work taught me something a simulation never teaches on its own: an answer is only correct if the physical thing it describes can be manufactured and will not overheat.
From there I moved into industrial instrumentation and machine vision — LabVIEW control systems synchronising conveyors, solenoids, PLCs and pressure switches on live production lines — then into medical imaging, building a GPU-accelerated OCT retinal analysis platform, and now automotive ADAS, where I lead AI-driven surround-view and voice-interface features on Renesas R-Car silicon.
Between July 2024 and August 2025 I worked on CUDA-EVAL with Turing, designing and reviewing GPU-programming evaluation tasks used to train and benchmark frontier language models. It is the same discipline as engineering design review applied to a new object: state the problem unambiguously, define inputs, constraints and acceptance criteria up front, produce a validated reference solution, and judge other engineers' submissions on technical correctness rather than plausibility.
The thread through all of it is modeling, simulation and validation: build a model of the thing, check it analytically, then check it against the real system — and be honest about the gap. I document as I go, because a result nobody else can reproduce is not a result.
Areas I have shipped production work in — not a list of things I have read about.
Transformer and reactor design to customer specification and IS/BS standards — magnetics, winding and core design, thermal and cooling calculation, insulation arrangement, and the production drawing set.
Building the model, defining operating conditions, and validating results analytically before trusting them — across MATLAB, LabVIEW instrumentation, robotic simulation and flight-simulation environments.
Schematic capture through layout and bring-up for 8/16/32-bit designs — high-current drive, sensor front ends, industrial buses, and the firmware that exercises them.
Closed-loop machine control on live production lines — motors, pneumatics, PLCs, pressure and proximity sensing — with multiple subsystems synchronised into one coherent data stream.
Profiling and rewriting the hot path — kernel design, memory coalescing, stream overlap, zero-copy CUDA/OpenGL interop — and getting neural networks and language models to run inside real latency and thermal budgets.
Problem definition with explicit assumptions and acceptance criteria, peer review of other engineers' work for correctness and completeness, and methodology documentation that lets someone else reproduce the result.
A year building what frontier labs are short of: hard, unambiguous, verifiable engineering problems with objective acceptance criteria.
A year on CUDA-EVAL, designing and reviewing GPU-programming evaluation tasks used to train and benchmark advanced language models. The hard part is not writing CUDA — it is writing a problem that has exactly one defensible answer, and a rubric that can tell a correct solution from a merely convincing one.
The other side of the same coin — not evaluating models, but making one survive contact with an embedded target.
Alongside my day role I design, build and operate a private production estate end to end — architecture, code, infrastructure, security and release engineering. Client and product names are withheld; the engineering is described below.
A five-node Linux server estate running more than twenty production web applications. Relay-based remote execution independent of inbound SSH, content-addressed resumable multi-gigabyte transfers with full SHA-256 verification, immutable releases and recorded rollback state for every deployment.
ERP with double-entry general ledger, CRM with workflow automation, file storage on S3-compatible object storage, real-time chat, video meetings over an SFU, webmail, time and attendance, and a shared identity service — one estate, many tenants.
A full identity provider: OIDC/OAuth2 authorization server with PKCE and RS256, a SAML 2.0 service provider hardened against XSW, XXE and replay, SCIM 2.0 auto-provisioning, and passwordless FIDO2/WebAuthn passkeys — built on the standard library rather than a vendor SDK.
Schema-per-tenant separation for the business applications and PostgreSQL row-level security for the shared ones, with the tenant GUC set at a single authentication chokepoint and FORCE row security so even the table owner cannot read across tenants.
Local LLM inference with per-organisation retrieval-augmented generation over pgvector embeddings, so tenant content is never sent to a third-party model API. Plus a cross-application copilot and rule-based assistants where a model is not warranted.
A browser-based 3D world with vehicle dynamics, physics, day/night simulation and multi-user presence — including draw-call reduction work, GLB asset pipelines, collision and surface-continuity audits, and GPU shader program validation against a hard frame-time budget.
CRDT co-editing over a WebSocket relay with scoped HMAC tokens, behind a suite of document, spreadsheet, presentation, PDF, forms, whiteboard and diagram editors sharing one portable engine core.
The same product surfaced as native applications across platforms — Kotlin and Jetpack Compose on Android, .NET 8 WPF and Qt 6 C++ on Windows, plus packaged desktop builds — with signed releases, in-app updates and CDN distribution.
Point-in-time recovery with WAL archiving and mutual TLS streaming standbys between hosts, encrypted off-host backups with validated catalogs, health probes, and documented failover procedure — designed so recovery is rehearsed, not improvised.
A cloud service that validates, stores and dispatches G-code to networked machines, and a six-axis motion controller on STM32H743 with grblHAL, a Wi-Fi co-processor link and Modbus spindle control — cloud and firmware on the same project.
Digital elevation models from LiDAR point clouds and from synthetic-aperture radar, GPU hillshade and colour-relief rendering for interactive terrain, custom QGIS tooling, a Windows port of a Linux GIS stack, and GPU-backed Kubernetes for the processing workloads.
Per-application credential isolation in systemd units, a WireGuard VPN for administrative access, automated TLS issuance and renewal across dozens of hostnames, least-privilege service users, and hands-on incident response including a live web-shell compromise.
Each one a problem, the approach, and what measurably changed as a result.
Customer-specification design of power, earthing and distribution transformers and reactors — the full path from electrical specification to drawings the production floor builds from.
Drivers get four fisheye camera feeds and have to infer where the car actually is. The feature stitches them into a single 3D scene the driver reads instantly, with obstacles raised as a vertical virtual wall so proximity is visible rather than inferred.
Optical Coherence Tomography scans the retina in cross-section. Clinicians needed the volume reconstructed, segmented and rendered while the patient was still in the chair.
Factory-floor vision and control with a hard constraint: the line does not stop, and the false-reject rate has to stay near zero.
Two robots at opposite ends of the scale, both solving a don't-send-a-person-in-there problem.
Power electronics to industrial vision to medical devices to automotive AI — each move added a layer rather than replacing one.
Leading development of AI-driven ADAS features — surround-view 3D visualization, obstacle rendering and an in-cabin natural-language voice interface — on Renesas R-Car hardware. Owns performance across the CUDA, OpenGL ES and machine-learning boundary, plus technical direction and mentoring for the feature team.
Designed and reviewed GPU-programming evaluation tasks used to train and benchmark frontier language models. Authored problem statements with explicit inputs, constraints and objective acceptance criteria, produced validated reference solutions, and peer-reviewed other engineers' submissions for technical correctness and completeness.
Pioneered the company's OCT imaging analysis platform: GPU-accelerated deep learning for retinal layer segmentation, real-time 3D reconstruction in OpenGL, and the Qt clinical application clinicians used day to day. Included two international onsite engagements.
Machine-vision inspection systems, industrial control integration and robotics — currency authentication, oil-seal inspection, a 600 kg material-handling robot and a bipedal radiation-survey robot for nuclear plant use.
Designed power, earthing and distribution transformers and reactors to customer spec up to 5 MVA / 33 kV — winding and core-cutting instructions, oil and cooling calculations, radiator and conservator selection, bills of material, and the AutoCAD drawings production built from, to IS and BS standards.
Built a real-time imaging pipeline end to end — capture, process, render, display — with CUDA processing in the capture engine, image registration in OpenCV/C++, and layer segmentation in MATLAB.
Designed MATLAB and C++ tooling for a medical imaging system to identify and characterise imaging parameters.
Self-assessed proficiency against real production use. The figure beside each is how long it has been in my working set.
Electrical engineering first, embedded systems second — which is why the hardware side never feels like someone else's problem.
Open to senior and lead roles in electrical, embedded and GPU engineering — and to contract work in simulation, modeling and AI training & evaluation. Comfortable with PST overlap.