Machine learning engineer, exploit developer in training
I build machine learning systems for a living — lately, mostly with large language models.
On my own time I’m learning exploit development on Windows x86-64, and writing the tools I need along the way.
Day job
This is most of my career and most of what I do at work today: building ML systems end to end, from data and features through training, evaluation and deployment. Lately, that means large language models.
On weekends, Jordan is usually
A toy: a hand-written probability tree, not a real model. Pick a token or sample.
Building on and evaluating today’s LLMs — where most of my recent work has gone.
Deep learning in PyTorch and TensorFlow, and classical ML with scikit-learn when it’s the right tool.
Feature engineering and reproducible training and evaluation pipelines. Measure first, then believe.
Models served from containerized FastAPI services, with CI/CD around them.
Most of my ML work is professional and not public. One open-source project from the computer-vision side:
Computer vision
Real-time YOLO detection on webcam or RTSP feeds, recording clips when something shows up.
Frames come in from a webcam or RTSP stream, detection runs on a background thread so the live feed never stutters, and short MP4 clips are saved whenever a detection fires. Any Ultralytics-trained YOLO weights drop straight in.
Python, YOLO, OpenCV, Flask
Side quest
Outside work, I’m putting serious hours into exploit development and vulnerability research, focused on Windows x86-64: reverse engineering, mitigations, ROP.
It started as a way to branch out into systems programming. Now it’s a goal: I want to be very good at this. Building tools is how I learn — writing a gadget finder teaches you more about gadgets than using one.
Exploit development on Windows
A ROP/JOP gadget finder that measures what gadgets do instead of guessing.
Most gadget tools infer a gadget’s effect from a hand-written instruction model. ropnroll runs every gadget on an emulated CPU with several input vectors and fits a closed-form relation — constant, copy, affine, bitwise — to the outputs. If the relation holds on every trial, it’s exact, and it’s correct for every architecture Unicorn supports, quirks included.
On top of that sits a best-first (A*) chain synthesizer that builds minimum-gadget call chains against the MS x64, cdecl and stdcall conventions, then verifies them under emulation before you ever touch a debugger.
Protocols and reverse engineering
RFC-style packet and memory-layout diagrams from a one-line field spec.
A pip-installable, cross-platform successor to the classic protocol tool. It ships with 20+ built-in definitions — TCP, IPv6, DNS, QUIC, Modbus, DNP3 — and takes custom layouts inline or as JSON.
The feature built for reverse engineering: paste a stack listing straight out of Ghidra and get a memory-layout diagram back, with ABI-implied slots filled in and pointers annotated, ready to drop into notes or an issue comment.
Python, CLI, Networking
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 ┌─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┐ │ Type │ Flags │ Length │ ├─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┤ │ Session ID │ ├─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┼─┤ ┊ Payload ┊ └─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┴─┘
A TypeScript port of pktfmt’s renderer, running in your browser. Output matches the CLI.
About
I’ve spent most of my career in machine learning, and I still do. Lately I’ve been pointing the same curiosity at the layer underneath.
My ML work has run from classical models and feature engineering through deep learning to today’s LLMs. It’s still my day job, and it left me with a habit of measuring things rather than assuming them.
Systems programming started as a way to branch out and keep learning. It turned into exploit development on Windows x86-64, which I’m working hard to get very good at. ropnroll is where the two meet: it fits relations to measured gadget behaviour instead of guessing it.
On GitHub
Pulled live from my profile, most recently pushed first.