Systems first, models second.
Agent systems
Systems where models propose and code decides: tasks with dependencies, memory, tools, policies, approvals and an audit trail. Jarvis is the working example.
AI-assisted engineering
I write the architecture, the rules and the acceptance criteria; coding agents such as Claude Code and Codex write much of the code under them; tests decide when something is done. Directing that work well is a skill in its own right.
ML and DL foundations
Python, NumPy, Pandas, introductory PyTorch and computer vision, alongside a computer engineering degree. Early days, and stated as such.
How my systems handle a request
Jarvis's request path, drawn on the hexagonal lattice of ice. Only one point is warm: the moment a human is asked.
- Transport
- Ingest
- Events
- Tasks + deps
- Proposed action
- Policy
- Approval
- Executor
- Audit + outbox
The honest stack
No percentages. Grouped by how I actually use each thing; levels in brackets are the ones on my resume.
Every day
- Directing AI coding agents (Claude Code, Codex, GPT)
- Product specs, architecture and acceptance criteria
- Git, Linux (Pop!_OS)
Built real projects with (AI-assisted)
- Next.js, React, TypeScript, Tailwind CSS
- SQLite, SQLAlchemy, FastAPI
- Playwright and pytest
- Model APIs via OpenRouter
Foundations I'm building
- Python (basic–intermediate)
- YOLOv8, object detection, ByteTrack tracking (basic)
- OpenCV, Ultralytics, NumPy, Pandas (basic)
- PyTorch (introductory)
Learning path
People detection, tracking & unique counting
Computer-vision learning project · Python, YOLOv8, ByteTrack, OpenCV, Pandas
An end-to-end pipeline with a pre-trained YOLOv8 model: annotated videos and CSV summaries of unique people and frame-level visibility. Built to understand data flow, inference and tracking, not as a production system.
A first end-to-end computer-vision pipeline: detection, tracking, unique counting.
Computer engineering coursework in programming; a real model layer inside Jarvis.
Data structures and algorithms; training models, not just running them.
Say something on the glass.
A website, an AI or software project, English lessons, or a job offer. Write a few lines about it; every message gets a personal reply.