These are the projects I've built, both on my own and as part of a team. Each entry explains what the software does, how it works and which parts I built myself.
voidwalk
Apr 2026 — now · personal
A binary analysis and disassembly toolkit for reverse engineers and security researchers. It detects whether an executable is a Linux (ELF) or Windows (PE) file, reads its sections and disassembles the machine code into x86 or x86-64 assembly.
C++20
Qt 6
CMake
GitHub Actions
The Qt 6 desktop app disassembling /bin/ls.
I wrote the x86 and x86-64 disassembler myself, with no disassembly library underneath. On /bin/ls, all 22,523 instructions it decodes start at the same addresses as in GNU objdump's output.
File formats and instruction sets are kept apart, so supporting a new architecture means writing one decoder class and registering it.
Files are memory-mapped, so a large program opens without being copied into memory first.
Decoding runs on a background thread. You can use it from a Qt 6 desktop app with syntax-highlighted assembly and a hex view, from a terminal UI, or from a scriptable command-line tool.
Every push runs 230+ tests on Linux and Windows through GitHub Actions.
Next is a debugger with breakpoints and tracing. Control-flow graphs, ARM decoding and AI-assisted analysis come after that.
An MCP server that lets an AI coding agent phone you when it finishes a long task or gets stuck, then hold a spoken back-and-forth with you. What you say goes back to the agent as text, so it can act on it. It works with Claude Code and other MCP clients.
Python
MCP
PJSIP
faster-whisper
Kokoro
GitHub Actions
Coding agentSIP over TLS · SRTP01:12
AgentUser
A call between the agent and the user (illustration only).
Each agent session runs a small stdio MCP server with four tools: call_start, call_say, call_end and call_status. One background daemon, reached over local HTTP with a token, owns the SIP registration, the phone line and the loaded speech models.
The daemon decides whether a call may happen, so an agent can't place one on its own. Calling stays off until it's switched on for a session, and there's also a global kill switch, snooze and weekly quiet hours. Declining a call snoozes calling for 30 minutes. No tool takes a phone number: the only number it can call comes from the config file.
Speech runs locally. faster-whisper turns your voice into text, Kokoro speaks the agent's replies, and voice activity detection finds the end of your turn. You can talk over the agent, and filler sounds cover its thinking time.
Calls are encrypted, with SIP over TLS, SRTP media and ICE/STUN. Linphone push notifications make the phone ring even when the app is closed.
Installers and a setup wizard cover Linux and Windows. On Windows the SIP daemon runs in WSL2, because the Windows build of the SIP library has no TLS. It ships as a Claude Code plugin, and the wizard can also write the config for Codex CLI, Claude Desktop, Cursor and Gemini CLI.
GitHub Actions lints the code and runs the tests on Linux, Windows and macOS for every push to main. A weekly run adds the tests that need the speech models and internet access.
A chat and file-sharing system for a local network: a C server on a single-threaded event loop, and a Python client with a PyQt6 desktop GUI. Clients find the server on the network by themselves, then join rooms, chat, see who's typing and share files.
C
POSIX sockets
Python
asyncio
PyQt6
The PyQt6 client on a live server, with three other clients connected. Solid lines are TCP, and the dashed one is UDP.
The server is one poll() loop over non-blocking POSIX sockets. It reassembles frames from partial reads, buffers partial writes until the socket can take more, and converts every integer to network byte order by hand.
Chat, rooms and message history travel as length-prefixed TCP frames, capped at 64 KiB. Discovery broadcasts, typing indicators and RTT pings use UDP, since losing one of those packets costs nothing.
Each upload or download opens its own TCP data connection, authenticated with a per-transfer token, so a large file doesn't hold up the chat.
Passwords are stored as salted SHA-512 hashes. The server ties a UDP endpoint to a session by the source address it actually sees, instead of trusting a port the client reports.
The client's asyncio protocol and network core has no GUI code in it. The PyQt6 interface sits on top, and can also run against an in-process mock server.
The C and Python codecs have to match the same byte-level test vectors. There are 130+ tests in total, including raw-socket tests and end-to-end runs against the live server.
A desktop app that gives civil engineers ready-made AI agents and lets them build new ones from a form, without writing prompts or code. My team of four built it in 48 hours at PoliHack v19 and placed 4th.
Python
Gemini API
PyQt6
The Egress Validator agent, which checks a parsed floor plan against the P118 fire safety rules.
My part was the engine that runs every agent, the Gemini client, the follow-up conversations and the DXF import. My teammates wrote the calculation tools and the desktop interface.
The engine runs an agent's tools first (the P118 rule checks, a Dijkstra search for evacuation distances), adds their results to the prompt, and parses the JSON in the model's reply into the agent's declared outputs, so the numbers in an answer come from the tools.
The Gemini client gives each request 30 seconds and retries rate limits and server errors with exponential backoff.
Agents keep the conversation for up to 10 follow-up questions. A keyword check sorts each question into on-topic, off-topic or unclear, and only the unclear ones go to an LLM classifier. A question that belongs to another agent gets pointed to that agent.
The DXF import turns closed polylines into rooms and corridors, by AutoCAD layer name or, when the layer isn't recognized, by shape. Door arcs become doors, a door that opens onto only one space counts as an exit, and each room's occupancy is estimated from its area. It also flags problems such as overlapping rooms or rooms without a door.
I wrote 33 tests for the engine, including runs of every agent against every example floor plan.