Apple Silicon ML — from research code to shipping product.
Open-source maintainer of ComfyUI-CoreMLSuite — one of the few
tools running Stable Diffusion through Core ML on the Apple Neural Engine,
1.5–2× faster on M2 Pro, maintained since 2023. Eight years of production
Python engineering before that.
You have a model that works. Now it needs to run fast, on-device, in
a real macOS or iOS app — or in a creative pipeline on Mac hardware.
I handle the Core ML and MLX conversion, the quantization decisions,
the attention implementation choices, and the Python or Swift glue
that holds it all together.
This is where my open-source work gives you the most leverage. Past
examples: 1.5–2× diffusion speedup via ANE optimization, batched
inference pipelines, custom node ecosystems for creative AI tooling.
My published benchmarks show the ANE running the SD1.5 UNet at
6–7× lower energy than GPU/MPS at the same speed —
the kind of measurement that decides where your inference bill goes.
Productized offer — performance audit. Benchmarks on
your hardware, a written report, and prioritized optimization paths.
1–2 weeks, fixed scope.
API design, data pipelines, deployment, observability — the
boring-but-essential work that makes a system you can actually rely
on. I've migrated enterprise codebases from Python 2 to 3, built a
market-making bot, designed CI tooling that outlived my time on the
team.
Technical advisory & code audits
When you need a second pair of eyes on a critical architectural
decision, a written review before committing to a direction, or an
honest assessment of a codebase you've inherited. Half-day to
multi-day engagements with a written deliverable.
Not sure where to start?
Book a free 30-minute discovery call →
— bring the problem, and I'll tell you whether I can help and what it
would take. If we don't fit, I'll say so and point you to someone who
does.
Selected work
ComfyUI-CoreMLSuite (open source)
A set of custom nodes that brings Apple Neural Engine acceleration to
ComfyUI workflows. 1.5–2× speedup on M2 Pro for
512×512 diffusion. Supports SDXL, LCM, ControlNet, LoRA baking, and
the full Stable Diffusion conversion pipeline. Used by the Apple
Silicon AI community since 2023.
Converting the VAE and CLIP to Core ML is near-lossless; the UNet carries essentially all the perceptual drift — so the full pipeline drifts no more than the UNet alone. Full ablation →LCM → Core ML conversion with ControlNet, sampled on the Apple Neural Engine.
blzr — a node-based editor for the Web Audio API.
Aims to cover the full API surface through a React Flow interface,
extended with utility nodes (MIDI), custom AudioWorklets, and
keyboard input. Built for fast prototyping of audio engines that run
entirely in the browser.
A seven-voice supersaw — detuned oscillators, stereo spread, dual filters, reverb — running entirely in the browser.
Alongside open source, I take contract engagements — usually as the
engineer who joins ML, software, and hardware components into one
working system.
Autonomous drone platform (contract, 2025–2026)
Python and integration lead on a multidisciplinary team (flight
control, computer vision, hardware). Built a low-latency
image-recognition pipeline on Raspberry Pi + Hailo
tuned for closed-loop flight control; designed training scenes in
NVIDIA Isaac Sim and a minimal custom simulator to validate control
logic before hardware deployment; drove YOLO fine-tuning for precise
feature localization. I was the one person responsible for making the
full software+hardware pipeline work end-to-end.
AI product delivery (via xfaang, 2024–2025)
Deployed and hosted an emotion-recognition-from-video
model behind a web application; built a voice-activity-detection
component for audio analysis; extended a Flask-based API with service
clients automating business workflows. Production AWS throughout
(Lambda, ECS, EC2).
From the log
I document the engineering — benchmarks, failed runs, exact versions —
at log.aszc.dev. Recent
findings:
An ablation pinpointing which converted component actually drifts, and
whether a human would ever see it (LPIPS).
About
I've spent eight years building software in places where it actually
has to work — enterprise networking at CodiLime, market-making at
Igoria, integration APIs and AI product delivery at Xfaang, edge ML on
an autonomous drone platform. I've migrated codebases from Python 2 to
3, mentored junior engineers, and built tooling that outlived my time on
the team.
In 2023 I started ComfyUI-CoreMLSuite because Apple Silicon ML
acceleration was a frontier no one had made accessible to
non-specialists. It's a project I still maintain — and the foundation of
how I work with clients today.
My instinct is for systems that are stuck — slow, brittle, or unclear
— and for figuring out which single lever actually moves them. That's
the work I want more of.
Remote, based in Poland (EU/CET). Working with clients internationally.
Contact
The shortest path: hi@aszc.dev
— I respond within one business day.
