Fraud Prioritization — Ranking transactions when only a limited number can be investigated.
|
MEMORY 01 / ACTIVE MASTER'S THESIS Fraud Prioritization is a master's thesis project about ranking credit-card transactions when investigation capacity is limited. Instead of treating every suspicious case equally, the system learns which transactions should be reviewed first under a fixed investigation budget. "There are receipts all over the desk. I only have room for a few in the tray... I hope I picked the right ones." PYTHON · LEARNING TO RANK · FRAUD DETECTION |
The tray represents the available investigation capacity. The receipts around it represent the larger pool of transactions that could potentially require attention.
The project studies how ranking models can prioritize cases under different budgets while keeping fraud probability, ranking relevance and financial-impact proxies conceptually separate. It places strong emphasis on provenance, verification and reproducibility.
Repository: Private
Life OS — A personal planning system for deciding what to do next.
|
MEMORY 02 / ACTIVE PRIVATE PROJECT Life OS is a personal planning and second-brain application that brings tasks, projects, goals, study, work, health and personal routines into one place. I built it to make the next useful action easier to see when everything feels equally important. "There are a lot of doors here. They all look important... I have been standing here for a while." WEB APPLICATION · PERSONAL PLANNING · SECOND BRAIN |
|
The doors represent study, work, health, personal projects and other responsibilities competing for attention at the same time.
Life OS brings those areas into one controlled system so that planning becomes a process of choosing rather than repeatedly remembering everything.
Less interface. More control.
Repository: Private
FH Agent — A perception, memory and planning architecture for autonomous RPG navigation.
|
MEMORY 03 / ACTIVE PRIVATE EXPERIMENT FH Agent is an experimental system for controlling a game through observed evidence rather than direct access to its internal state. A local language model acts as the planning layer, while separate components handle perception, memory and safe input. "We gave the robot a map so it would stop getting lost. It is still getting lost... just more thoughtfully now." PYTHON · AGENT ARCHITECTURE · LOCAL LLM · SAFETY |
The image captures the difference between having information and reacting to it correctly.
The agent may understand part of its surroundings while still missing an immediate danger. This is why the architecture treats perception, evidence, memory, planning and execution as separate responsibilities. Early milestones focus on observation, logging, no-spoiler boundaries and controlled actions before adding long-running autonomy or reinforcement learning.
Repository: Private
FireRed Randomizer — A compatibility workspace for a modified FireRed-based game and modern randomizer tools.
|
MEMORY 04 / ACTIVE PRIVATE WORKSPACE FireRed Randomizer is a documented workspace for combining a heavily modified FireRed-based game with Generation 9 content, a modern randomizer and supporting emulator tools. It records compatible versions, build steps, runtime checks and known limitations so that the setup can be reproduced reliably. "A strange creature appeared during testing today. I do not remember putting it there... but it seems happy to be included." JAVA · GAME MODDING · COMPATIBILITY ENGINEERING · TESTING |
|
The creature represents unexpected combinations and visual edge cases that appear when older game systems are extended with newer data and randomized rules.
The workspace turns those surprises into documented policies, repeatable builds and targeted runtime tests. ROMs, saves and other protected artifacts remain outside version control.
Related public repository:
Universal Pokémon Randomizer FVX
I'm using Hyperskill as a project-based way to deepen my programming knowledge.
Rather than treating progress as a checklist, I keep a growing Knowledge Atlas of the concepts behind the projects I build. It keeps the full learning landscape visible at once — including topics, prerequisite and dependent relationships, learning progress, and the project evidence connected to each part of the map.
As I complete projects or verify new topics, the atlas evolves with the underlying learning data while preserving its existing structure.
Current course: Introduction to Java
Current progress: 31 / 89 course topics learned · 12 verified
Explore the interactive Knowledge Atlas →
Browse the Hyperskill project repository →
Hyperskill Projects — practical evidence behind the atlas
| Project | What I built | Atlas evidence |
|---|---|---|
| Simple Chat Bot with Java COMPLETED · HYPERSKILL |
A five-stage console chatbot that introduces itself, reads user input, calculates an age from remainders, counts with the user and finishes with a small quiz. | 26 required topics connected back to the knowledge hierarchy through project coverage. |
DATA AND MACHINE LEARNING
SOFTWARE AND SYSTEMS
PROCESS AND ARCHITECTURE
The typing animation is provided by Readme Typing SVG .