I design systems for constrained compute
The hardware changes — Jetson on a drone, Apple Silicon on a desk, ARM in a rack. The constraints stay: limited power, limited memory, no guaranteed connectivity. Constraints force clarity. I design systems that work anyway.
System design under constraints
Constraints force better design. When you can't throw hardware at the problem, the architecture has to be right. I design systems for environments where every watt, every byte, and every second of latency matters.
Autonomous BVLOS Drone Operations
500+ hours flying 32kg+ eVTOL drones beyond visual line of sight. 200km+ single-leg flights across African savannahs and the Amazon rainforest. Point-to-point charging enables unlimited mission range. Multi-country regulatory experience.
Cloud Infrastructure & Compute
Edge-to-cloud architecture for autonomous systems. Telemetry pipelines over satellite links, real-time data processing, microservice APIs. Cloudflare Workers at the edge, FastAPI backends, containerized deployments on constrained hardware.
Edge Computing & Embedded Systems
Built the Jetson software stack running onboard a drone at 100km/h over the Amazon. SSH into an Nvidia computer mid-flight over Starlink to debug and run diagnostics. Sensor integration, hardware-software coupling, real-time processing under power, thermal, and bandwidth constraints.
Machine Learning at the Edge
ML inference on constrained hardware. Designed the cloud pipeline for autonomous data collection feeding ML models. CUDA-accelerated processing on Jetson. The full loop: collect data in the field, train in the cloud, deploy back to the edge.
Local AI Training on Apple Silicon
Training ML models locally on M-series chips. Private by design — data never leaves the device. Fine-tuning LLMs, training custom models overnight. Using MLX, PyTorch MPS, and the Apple Neural Engine. Same edge compute philosophy, different hardware.
Space Technology
Satellite tracking and orbital mechanics. Conjunction screening for 30,000+ objects. The drone-to-space transition is natural. Same problems at different altitude: constrained hardware, degraded links, can't touch it once it's up.
Things I've shipped
Open-source tools, real-time applications, and systems that run 24/7 on a $250 computer.
SatTracker.
Real-time 3D satellite tracker. 30,000+ objects rendered on a WebGL globe with live orbital propagation. SGP4 in the browser, updated daily from Space-Track.org.
sattrack.io →OrbVeil.
Satellite conjunction screening. Screens 30,000+ objects daily for collision threats in under 10 seconds. Open-source Python library with automated CDM analysis.
orbveil.com →Arsenal.
Suite of 7 internal tools built in 7 days. Content pipeline, document vault, design system, task management, system dashboard — all self-hosted on a Jetson Orin Nano.
autmori.com →Gale Cloud Console.
Flight analysis platform for autonomous eVTOL operations. ArduPilot log parsing, energy efficiency metrics, multi-mission fleet analytics. Built for 19Labs.
PrivateRustane.
Training, inference, fine-tuning, and distributed compute suite built in Rust for Apple Silicon. Direct Apple Neural Engine access via private APIs — zero GPU, zero cloud.
GitHub →The path
19Labs
Building and operating a 32kg eVTOL drone with Starlink Mini satcom, 200km+ single-leg range, fully autonomous BVLOS. Thousands of km flown across African savannahs and the Amazon rainforest. Built the Jetson software stack that runs onboard. Designed the cloud data pipeline for autonomous data collection feeding ML models. SSH into the aircraft mid-flight for debugging and diagnostics.
AI Development
Working with AI since early 2023 — before ChatGPT was mainstream. Built various applications leveraging LLMs for automation, content, and system orchestration. Currently using voice-driven AI development to ship 7 production applications in 7 days on a $250 computer.
Autmori
Open-source space intelligence. OrbVeil screens 30,000+ objects daily for collision threats. SatTracker renders every active satellite in real-time 3D. The drone-to-satellite transition is natural — same constrained hardware, same degraded links, same "can't touch it once it's up" problem.
Let's build something
Available for consulting on edge AI, local model training, autonomous systems, and space technology. I work best on hard problems with real constraints.
I hijacked Apple's Neural Engine — the chip built for Siri and photo filters. Reverse-engineered the private APIs and trained a full LLM on it. Zero fan noise. Zero GPU. Just the Neural Engine doing what nobody thought it could.
I blew up another model on my MacBook Pro. I'm pretraining a 48.8M param GPT in Rust on Apple's Neural Engine. This is my first Rust project.
I trained a GPT on my MacBook Pro. No cloud. No NVIDIA. Just the laptop on my desk. 459 lines of Objective-C talking directly to Apple's Neural Engine.
Porting ANE training from Obj-C to Rust. Here's what I validated first.