One
engineer,
whole stack

Mechanical Design & Product Engineering

For hardware that has to exist physically, work repeatedly, and be manufacturable

Mechanism and assembly design from requirement to released drawings: kinematics, load paths, materials selection, tolerance stacks, fits, and design for manufacture. The underlying coursework is a complete mechanical engineering degree — machine design, thermodynamics, fluid mechanics, dynamics, strength of materials, manufacturing processes — and the practice is a decade of building things afterward, from a 130-part mountain board designed from scratch to automated-storage installation logistics for a $1B distributor.

• Concept through detailed design, with the trade-offs written down rather than implied
• Mechanism design, load and thermal analysis, and failure-mode review
• Materials and process selection, DFM review, tolerance stacks
• Bills of materials, sourcing, and vendor-ready drawing packages
• Prototype iteration — 3D printed, machined, or laser-cut

Electrical & Embedded Systems

For devices that run unattended, on a budget of power and bandwidth

Microcontroller systems end to end: firmware, protocol, power, and the regulatory reading that quietly determines what you are permitted to build. Two products in the portfolio exist largely as a protocol argument settled correctly — one puts the intelligence in the panel because the Bluetooth link carries 1–4 KB/s, the other puts it in the host because USB does not. Same family of product, opposite answer, written down with the reasoning.

• ESP32 and Arduino firmware; C and C++; PlatformIO and ESP-IDF
• BLE peripherals, custom GATT services, and serial framing protocols
• LED driving at panel scale (FastLED, WS2812B), timing, and refresh budgets
• Power budgeting, including automotive supply and weatherproofing constraints
• Sensors, actuation, and the host-versus-device intelligence split
• Regulatory briefs — what the rules force on the design, before you build it

CAD, 3D & Architectural Modeling

For anything that has to be drawn accurately before it can be made or sold

Extensive parametric CAD across mechanical and architectural work — five years in SolidWorks, two in Onshape, plus the graphical-communication, materials and manufacturing coursework that makes a drawing mean something to a machinist. Modeling for manufacture and modeling for visualization are different jobs, and this practice does both.

• Part and assembly modeling, including large assemblies and configurations
• Manufacturing drawings, GD&T, and revision-controlled release packages
• Architectural and interior 3D modeling for visualization and planning
• 3D-print, CNC, and laser-cut preparation with the process constraints applied
• Reverse-engineering an existing part into a clean parametric model
• Renders and exploded views for documentation, instructions, and sales material

Machine Learning & Applied AI

For models that must work on your data, on your hardware, verifiably

Architectures built and trained from first principles rather than imported from a tutorial: variational autoencoders, convolutional encoders and decoders, diffusion models, neural style transfer, and learned controllers. Published work puts machine learning in the loop of a physical robot; shipped work trains a VAE on an artist's own paintings in three minutes on a laptop.

The habit that matters most here is verification. In Pixelwall, the browser's JavaScript decoder is checked numerically against the PyTorch reference — same weights, same latent, agreement to within uint8 rounding — because hand-indexed weights make it very easy to ship a plausible picture that is not the trained model. That discipline is the deliverable as much as the model is.

• Problem framing: whether this is even a learning problem, and what the baseline is
• Dataset construction, labeling strategy, and honest train/validation discipline
• Model design and training in PyTorch; checkpoint selection by validation, not by last epoch
• Evaluation that would survive a reviewer, not just a demo
• Deployment — including on-device and fully offline inference
• Porting a trained model into a production runtime, with numerical parity proven

LLM & Agent Systems

For language-model projects that have to leave the demo

Most LLM work stalls in the same place: it demos beautifully and cannot be trusted in front of a customer. The fix is unglamorous and mostly architectural — rules enforced in code rather than in a prompt, structured outputs with schema validation, evaluation harnesses that catch a regression before a user does, and a defined fallback for when the model or the network is unavailable.

Both shipped examples do this literally. One game's referee overrules the model when it breaks the rules and ships a complete offline brain so the product works with zero network. One research system runs a fifteen-persona boardroom whose grades are scored against reality every evening and fed to a neural net that learns which personas to trust.

• Multi-agent and multi-persona architectures with a defined arbiter
• Structured JSON output with validation — downstream code can rely on the shape
• Retrieval over your own documents, with citations back to the source
• Evaluation harnesses and regression suites for prompts and models
• Cost and token control: caching, context trimming, routing cheap work to cheap models
• Graceful degradation — an offline path that keeps the product usable
• Human approval gates wherever the action spends money or touches a customer

Robotics, Control & Simulation

For motion problems where the standard approach has already failed

The published specialty. Limbless and soft robotic locomotion, pneumatic actuation, real-time feedback control derived from partial differential equations rather than a lumped approximation, geometric mechanics and variational integrators, and machine-learning-driven burrowing in granular media — presented at IEEE ICRA, IROS, and ISER.

If a mechanism has to move through something difficult — sand, soil, a pipe, a body, a confined channel — this is not an adjacent skill. It is the exact problem, and there is a peer-reviewed paper on it.

• Locomotion design and gait optimization, including in granular media
• Soft and compliant actuator design, modeling, and characterization
• Control design: PDE-based, model-predictive, and learned controllers
• Simulation that matches the bench — and the experimental protocol to prove it
• Experimental rigs, instrumentation, and data collection
• Feasibility studies and technical review before you commit a budget

Software, Web & Mobile

For the layer your users actually touch — designed as well as built

Full delivery across an unusually wide platform range, all of it represented by something finished: edge compute on Cloudflare Workers and Durable Objects, React Native and Flutter on mobile, SwiftUI natively, Unity for real-time 3D, and dependency-free web applications that install by opening a file.

Design is included. Every interface and every site in this portfolio — including this one — was designed and built by the same person. You are not paying for a developer who then needs a designer, or a designer who then needs a developer.

• Web applications, marketing sites, and interactive tools
• iOS and Android — React Native, Flutter, or native Swift
• Edge and serverless backends with genuine state, not just functions
• Desktop packaging for macOS, Windows, and Linux
• Interface and visual design, typography, and design systems
• Technical writing: specifications, protocol docs, and documentation people read