Hun Labs

Labs

Projects I have built with AI coding agents since the summer of 2025. I started them as my interests led, so the fields vary, but the way of working runs through all of them. The story is in Story.

AI coding agents wrote most of the code and documents. I decide what to build and why, and I check what is right. Status and dates are given by month.

The project I put the most into right now — TTAgentRing · a round touch device on my desk that shows my AI coding agents' state as ring color (hobby · in development)

How the work is built

architectures

A documentation hub that maps every repo in one visual language

Active · last activity 2026-10
  • Generates and maintains architecture docs and diagrams for all of my personal repos
  • One canonical set per project: document, graph source and rendered diagram
  • Shared color and line rules make diagrams comparable across repos at a glance
  • Detects changes in source repos and checks whether the docs are still fresh
  • Extracting the AI coding agent setup into a common scaffold (preparing for release)

TTClickCell

AI-assisted Excel automation desktop app — my first AI coding project (ended)

Ended · last development 2025-10
  • Filters, de-duplication and conditional formatting through buttons and natural-language commands
  • An Excel engine that keeps the original formatting when saving after edits
  • The first project where I gave a coding agent a working charter and a change-log protocol
  • Stopped by choice: while my day job and research came first, Copilot in Excel matured and the window I was aiming for closed
  • Starting point for how all later repos are built and for the CKM research — the repo is kept

Research

cognitive-os-paper

Research on structured prompting and LLM behavioral consistency

Active · last activity 2026-09
  • Empirical study of how structured state records affect the consistency of LLM responses
  • Consistency metrics quantified from large-scale observations across many LLM models
  • Experiments, analysis, writing and audits managed in a single pipeline
  • arXiv preprint (arXiv:2607.24765, not yet peer reviewed) · the reproduction package went public on Zenodo first

cogos-behavioral-consistency

The public reproduction package for the LLM behavioral consistency paper

Done, in use · last activity 2026-07
  • Code and data bundled so a third party can recompute the paper's numbers independently
  • Results reproduce from the bundled data alone, with no new API calls
  • Includes integrity checks, with separate licenses for code, data and paper
  • Released publicly alongside the paper

AI and work tools

ttpatent-mcp

A tool for looking up public patent, trademark and design data from an AI chat

Active · last activity 2026-09
  • Search domestic and international patent, trademark and design records right inside a conversational AI app
  • Read-only by design, with no write or save functions at all
  • Defensive structure that surfaces empty results as errors instead of passing over them silently
  • Explicitly refuses legal conclusions such as registrability or infringement
  • In operation

trademark-automation

An assistant for trademark clearance searches and goods/services suggestions

Active · last activity 2026-09
  • Enter a mark or product name to search for similar prior marks in public trademark data
  • Risk graded by deterministic rules centered on the distinctive part of the mark
  • Approval gates: nothing moves to the next step without an expert's sign-off
  • Local-first design that never sends case data outside
  • Results are reference material only and never replace legal judgment
  • Built at the request of a patent attorney I know · core flow done, refinement in development

ttbookforge

A publishing pipeline that turns varied source documents into e-books and educational comics

Active · last activity 2026-09
  • Converts very different sources — classical texts, rulebooks, scanned PDFs, classical Chinese records — into e-books
  • Translation, proofreading and typesetting run as separate pipelines per track
  • Includes tracks for digitizing scanned books and producing an educational comic encyclopedia
  • A parallel track writes original technical books from scratch
  • Orchestrated by AI coding agents, with progress handed over between sessions
  • Development and production in progress per track

offllm

A local LLM inference server for an internal network, with no external cloud

Dormant · last activity 2026-04
  • Runs large language model inference locally on a single desktop GPU
  • A standard-compatible API, so existing code works by changing only the address
  • Swaps models in only when needed, within memory limits
  • Firewall setup that allows internal-network access only, plus health-check scripts

PaperPipeline

An AI pipeline for writing and checking papers, tailored to a research lab

Dormant · last activity 2026-04
  • An AI coding agent workspace for paper writing, packaged for reuse
  • Cross-checks numbers in the paper against the source data
  • Traces claim–evidence–data links and simulates reviewer attack surfaces
  • Audit agents for experimental design, statistics and data integrity
  • A personal project built to fit a research lab

latexcp

A one-click builder that turns Korean LaTeX documents into PDF

Done, in use · last activity 2026-05
  • Desktop GUI that converts Korean LaTeX and Markdown documents to PDF
  • Injects font, margin and paper settings automatically to minimize Korean typesetting errors
  • Built-in editor and build history tab
  • Ships as a single executable that runs without installation

ttsoft-tts

A Korean speech synthesis and transcription studio running on a local GPU

Done, in use · last activity 2026-06
  • Speech synthesis that turns text into natural Korean speech
  • Transcription of recordings with speaker separation
  • A single-slot setup that swaps synthesis and recognition models on one graphics card
  • A terminology dictionary step that reduces misheard technical terms
  • Operated within the terms of the model licenses

Hardware and hobbies

ttcalslate

An engineering calculator slate on an ESP32 touch board

Active · last activity 2026-09
  • Numeric, unit-aware and symbolic modes on one calculation engine
  • Designed so one mode can pick up the result of another
  • Arithmetic correctness checked against an independent implementation
  • Verification of computation kept separate from verification of display and input
  • On-board testing in progress; in development

harvis

A hobby indoor autonomous four-wheel rover, unit one

Active · last activity 2026-09
  • Aims for a loop in which an AI agent on a home server observes, diagnoses and fixes the rover
  • Schematics, layout and assembly documents produced by generators
  • "The hardware beats the docs": every verification claim carries a grade of physical evidence
  • Drive train confirmed running on the real unit; power and firmware in development

teos

A self-built measurement kiosk OS on an ESP32 board

Active · last activity 2026-09
  • Aims to be a logging hub that records the whole experiment, not just values
  • Multi-channel sensor measurement with SD-card logging from a touch-screen UI
  • Sensor drivers written in-house for thermocouples, temperature/humidity and analog inputs
  • A built-in camera photographs analog gauges on a schedule, aligned in time with the data
  • Every "it works" claim carries an on-device evidence grade, verified against reference values
  • A Linux-board prototype ported down to a microcontroller in one piece

ttslicet3d

A 3D-print output pipeline built on an open toolchain

Active · last activity 2026-09
  • From slicing to real prints end to end on an open Linux toolchain
  • A safety audit that enforces temperature and build-volume limits in one place right before printing
  • Print-time estimates based on motion planning
  • A mesh toolkit — mirroring, plane cuts and search for the orientation that needs the least support
  • Works with sibling projects through file contracts that machines can check

asterism

A near-future sci-fi tactics RPG developed autonomously by AI coding agents

Active · last activity 2026-10
  • Turn-based strategy RPG on a 2D grid, with three factions clashing along different tactical axes
  • World, rules and code structure deduced from a small set of axioms
  • Faction differences verified by automated tests on both numbers and actual behavior
  • An autonomous branch with no human input was run as an experiment, then merged into the main line
  • A design loop that tunes difficulty with batch simulations; in development

terrain-globe

Printable terrain pieces generated from public elevation data

Dormant · last activity 2026-08
  • Combines public land elevation and seafloor data into 3D terrain meshes
  • A general generator: give it a coordinate region and a scale
  • Produces watertight pieces that print without supports
  • Redesigned from the projection up so seams line up, confirmed by fitting printed pieces together
  • A large modular globe (cube-sphere split) at the design stage
  • Started as geography and geology teaching aids for kids

tlcn

A control node that switches one living-room TV between media and retro games

Dormant · last activity 2026-05
  • Detects media playback and game launches and switches the TV input automatically
  • Device-to-device control signals first, falling back to infrared when they fail
  • A small screen on the board shows the current mode and status
  • Designed so a physical button can always switch manually even if automation stops
  • Features complete; at the acceptance test stage

slimblade-config

A trackball settings GUI for the Linux desktop

Dormant · last activity 2026-04
  • Visual editing of button remapping, natural scrolling and cursor speed
  • Rule editor that maps simultaneous button presses to shortcuts
  • Updates only its own block while preserving the rest of the existing config file
  • Settings persist across reboots

Other tracks

  • A casual chatbot running open models on a local GPU
  • A finite-element structural analysis design loop
  • LLM autonomous-judgment experiments · shadow records
  • Open-source CFD reproduction and verification
  • A web dashboard for remote air-conditioner control and indoor temperature/humidity monitoring
  • Design exploration on a predict–measure–residual cycle
  • A reasoning workbench that supports decisions through consensus among several LLMs

Tracks listed by field only because they touch on work, reviews or security.