Frank's Projects

02 / Projects

Things I've Built

Projects I've built in my free time. Each one has taught me something new.

Computer Vision

CatCam

A motion-triggered camera on a Raspberry Pi watches my pet door and saves every event to a NAS. It also captures the motion of sunlight and shadows, so most of the tens of thousands of events contain no cat at all. I built a pipeline that triages each event with YOLO and escalates only the ambiguous ones to a local vision model through Ollama, which keeps the expensive calls to a small fraction.

I hand-labeled a confirmed dataset with a custom hotkey sorter, then used it to tune the detection threshold and pick the smallest vision model that didn't mistake shadows for cats. Every model change gets scored against that dataset to compare accuracy.

Python YOLO Ollama + VLM OpenCV

Reverse Engineering

Private Server Emulator

A private server created from scratch for the MMO Dofus, written in Python. It uses the same network protocol that the stock, unmodified client does and runs a server-authoritative world. It recreates the combat, monster AI, and economy, in addition to some custom content on top.

It started as a MITM packet sniffer. Once I had generated the full message codec from the decompiled client, I iterated on it feature by feature until it was fully playable offline!

Python Packet Interception Protocol Codec asyncio

Full Stack

CenoColors

An unofficial game mod that lets players pick custom colors for in-game animals. A Flask site collects the player's choices, the app rewrites the color hex values in the game files, and a Discord bot handles the injection over SFTP. This is the project that taught me reverse engineering.

Flask SFTP Discord Bot Hexcode Modification

Try the Animal Builder at CenoColors.com

Data Pipeline

Resource Spotter

A computer vision model that identifies resource types in Albion Online. At its core is YOLOv11, fine-tuned on a dataset I collected and labeled myself.

Behind it is a full data pipeline, including automatic labeling for new data, so the model keeps improving without me hand-labeling every frame.

Python YOLOv11 Computer Vision

AI Training

Agent Snek

My hands-on lesson in reinforcement learning, end to end. I designed and trained an agent that plays Slither.io on its own, with Selenium driving the browser and an LSTM + PPO policy driving the snake.

Python Reinforcement Learning Selenium LSTM + PPO

03 / Background

How I Got Here

Spreadsheets to Python

I've always built spreadsheets to track whatever data interests me. Once I realized I could act on data through automation instead of only tracking it, I learned Python to do exactly that. Since then, bots can do far more than I thought and my entire worldview has changed because of it!

I'm a project learner. When I want to pick up a new technology, I find a personal use case for it, usually something with games, data, or automation, and start tinkering.

When I wanted to understand computer vision, I spent weeks collecting and labeling my own dataset, then trained a model to identify resources in-game. I could have grabbed a dataset from Kaggle and followed a tutorial, but then the project wouldn't have been mine.

Python to AI

These days most of my building happens with agentic coding. I define the architecture and write the design specs, and I keep a mental model of what's under the hood. I set strict completion criteria up front and run verification loops to catch issues before PRs land, which means less slop to rework. I still decide what features get built and how the project is structured.

What I've Learned

Building things has taught me far more than whatever I initially set out to learn. I'm proud of these projects because each one is a stepping stone, and what I learn in one shows up in the others.

I like to build with interesting tech