Chang Hyun WooAI engineer, LLM agentsSeoul
I design each system and decide what it has to do before it is built. Then I test it, find where it is weak, fix that, and test again until it meets the standard I set. Scroll to see the systems, how each one works, and the parts you can try yourself.
Index
Each one links to how it works and what I measured.
An AI VTuber that talks back
She hears you, decides whether you were talking to her, and answers in her own voice with a moving avatar.
Shared memory for coding agents
One agent writes down a status or raises a blocker. The next agent, cold, in another repo, reads it and picks up the work.
Two games you can play on this page
Lumina Raid, multiplayer flight combat, and Rhythm Blaster, whose charts are built from the song itself.
An art style I trained, handed to a whole community
A game master for voice Mafia
A vocal out of a mix, when you own the instrumental
01Profile
Every number on this page has a note behind it. The ones that went against me are here too.
Before any code exists I write down what the system must do and how we will know it does: the cases, the pass mark, and what would count as failing.
After planning, I write the code. I read it, correct it and send hostile reviewers after it. The corrections are the record of what I actually did.
Held-out cases, blind labels, paired tests. When one of my own numbers turns out wrong, I retract it where I published it and fix the cause. Two of those are further down this page.
02Lumina
Discord, Telegram and spx6900.com. Voice, text, vision, memory, games.
Lumina is a character people talk to in a voice call. She hears you, decides whether you were talking to her, answers in her own voice with a moving avatar, remembers you, and plays games on stream. Her personality is a language model I fine-tuned and run on my own GPU; small, fast classifiers around it make the yes-or-no decisions. What matters to the person talking to her is how quickly she answers.
Discord voice arrives encrypted per speaker. She waits half a second to be sure you have stopped, then turns your speech into text in about a fifth of a second.
Her three nearest memories about the speaker come from a local vector store in about a tenth of a second. I removed a one-second cloud call that only decided whether to look.
Two things start at once: a quarter-second check decides whether she was being spoken to, and her reply starts being written. She speaks only when both are done, so the slower one sets the pace.
Her voice starts a quarter of a second later, streaming the first sentence while the rest is still being written.
The same audio plays on my machine, and that local playback drives her mouth, so lips match what you hear. A last quick check picks her expression.
Try it
This replays one turn at real speed using the medians I measured on 21 Sep 2026. Switch between the old gate and the new one and run it again: with the old gate, the gate was the slower branch in 18 of 20 turns.
Try it
For every line said in the call, the gate returns a probability that she should speak. Drag the cut-off and watch what happens across 108 real voice moments she had never been tested on. These are the recorded results, not a simulation. Missing someone who spoke to her is the error that matters; butting in is cheap.
Example lines were written by hand for testing. The 108 test moments came from real calls; only their scores are on this page, never what anyone said.
One gate design scored 95% after I adjusted it against its mistakes. So I froze every design, fingerprinted it, had 208 fresh cases picked and labelled blind, and ran again. It fell to 86.4%. A simpler design held up at 92.9%, and that is the one that shipped.
The fine-tuned model wrote its stop marker as plain text instead of the special token, so generation never stopped early and every text reply spent its whole 2,000-token budget. It had been blamed on a search feature shipped the same week. I added a check to the local server that stops as soon as the marker appears as text: a test reply with the same 2,000-token limit ended at 56 tokens, in 12.5 seconds. I also trained later versions on the real stop token.
Recall returned the right memory first, at cosine distance 0.195. It arrived as a tool message that the local server's formatter silently dropped, so the model never saw it. Retrieval was never the problem. I changed the bot to send recalled memories as plain text inside the user's message, which the formatter keeps. By 18 Sep she was repeating stored facts in her replies.
Lumi decides what to do and plan it out throughout the run and uses a harness to execute the move she planned to do. While she plays, she does live commentary, and chooses the recovery when she gets stuck. No human intervention, all ran locally on my PC. The recording is 1 h 46 min, unedited, 31 Aug 2026.
03OpenVault
48 MCP tools · 193 tests · MIT. Live since July 2026.
One agent writes down a status or raises a blocker. The next agent, cold, on another machine in another repo, reads it and picks up the work. Nobody relays the message. It is a Next.js app and an MCP server with a knowledge graph, a mirror of the code, and a review queue a member agent cannot approve its way around.
An agent that has never seen the project asks the vault where things stand. It gets a short summary built from real notes, about 155 tokens when measured, with any proposals waiting for review.
The vault keeps a copy of the code, synced from git. The agent reads the real file from that copy. It needs no clone and no access to the repository.
The agent cannot push, so it files the change as exact before and after text, with a written reason. The reason is saved as a note that outlives the review.
Only an account with the owner or executive role can approve, and approving changes no code. The owner's agent takes the approved result, writes it into the owner's own checkout, and pushes.
The push starts a workflow that copies the new commit into the vault. The suggestion then shows as applied. The vault never writes to git.
Try it
This replays one change from an agent that cannot push, across three lanes: the agent, the vault and the owner. Click any step, or walk through with Next, to see the real tool call and what comes back. Run plays them all and can be paused. Nobody timed these steps, so the replay shows no durations.
Four questions, three mechanically checked must-mention terms each, scoring token cost and answer coverage together. Coverage exists because cost alone scores a cheap wrong answer as a win. Code explains how a feature works better than a note can, so I kept that loss and wrote the lesson into the README: notes hold decisions and rejected options. A later review caught grep reading the README, which reprints the answers; I excluded it, re-ran, and got 2 of 4 again.
I had published "33x cheaper". The code map it measured is a manifest with no source text, so the ratio compared "which files changed" with "what does the code say". Re-measured, the reading saving is near zero. I replaced the ratio with plain token counts and added a rule to the instructions the vault gives agents: a ratio between two different questions is not a saving.
There are two review gates and a member is refused at both. A member agent that can push still cannot mark its own work done. An owner or executive has to approve it first, and that check is one shared implementation behind both the MCP tool and the web route, so the rule cannot drift between them. An agent that cannot push files a suggestion instead, as in the replay above. Either way the vault stores no git credential. Whoever pushes does it from their own machine.
Found with a two-arm experiment, not by reading docs: a session-start matcher of "startup" never fires on resumed sessions, exactly the ones that most need a briefing. Every resumed session had silently received nothing. I set the hook to run on all four kinds of session start and to print the reason when a briefing fails. Over the next five days I counted 23 briefings delivered on resume in one project's logs.
04spx6900.com
The official SPX6900 site. React on Vercel. 8 sections, 2 games, a meme maker.
Try it
Three.js on the client, one Cloudflare Durable Object as the shared sector. Your browser owns your movement so it stays responsive at distance; the room owns hits and score, and caps what any single hit can claim at 160. If someone else is online you will see them.
New · 25 Sep 2026 I rebuilt all six ships as faceted models, one file per ship, and redrew the flight HUD as thin instruments. A dev-only hangar renders each ship headless, so a change is judged from a screenshot.
Try itRhythm Blaster
An aim game with no button to press: squares blast out on the beat and you score by having your cursor where each one lands. The chart is not hand-made. Your browser listens to the song, finds the beats and builds the same chart every time. This is one track from the full game, sung by Lumina.
The room's id counter reset whenever Cloudflare evicted it, so every pilot was issued id 1. Every scripted test passed. It only showed up with two real browser tabs. Multi-client behaviour is tested with real clients now.
Onset detection runs in your browser and produces the same chart every time. 14 songs, 3 difficulties, 42 public leaderboards. A flawless run scores exactly 1,000,000.
PlayA layered editor where preview and export share one render function, so what you see is what you save. The cut-out model (U-2-Net, about 42 MB) runs in the browser; your photo never leaves your machine.
Open05More systems
An art bot, a game master and an audio tool.
Artists send a photo or an idea to a Telegram bot and get it back in the Project Aeon style, the official SPX6900 NFT collection. I taught that style to an open image model from 26 examples: FLUX.1 first, moved to FLUX.2 in December 2025. It runs on my own GPU, so nobody pays per image. Inspired by Milady Station.
The one call that ever broke was the captioner: an upstream model swap started returning refusals as a normal HTTP 200, and they were being drawn as art. Three failure signatures now each get their own handling.
Hosts Mafia in a voice call for 5 to 12 people: mutes players outside their turn, hands out roles privately, runs night actions, trials and votes. Lumina can join as a player from her own bot; her "was that said to me or about me" gate scored 23 of 24.
Public repoNot blind separation: it works out how the instrumental was shifted, drifted, EQ'd and limited inside the mix, applies that, and subtracts. On a mix with an offset, a gain change, inverted polarity and EQ, plain subtraction scores −3.9 dB; this scores 32 dB. About 10 seconds for a 2.5 minute song.
06Record
A community internet-culture project. I own the agent and ML stack end to end: Lumina, her fine-tuned persona model, the image bot, the game-playing agents and the site. Not the token, trading, or any financial product.
The shared-memory server for coding agents described above.
Private investment firm of SM Entertainment founder Soo Man Lee. Led diligence on a $750K graphene manufacturing investment that returned 52.6% within four months.
ROK engineering contingent. Worked across Arabic, English and Korean for an 8-nation engineering mission: 114 UN Engineering Task Orders and 108 km cleared of mines. Four decorations including the UN Medal.
Intern broker on the interest-rate swap desk; intern copywriter, localising a site into English.
Contact
