●Software engineerAI systemsMalaysia
I build software that checks its own work: AI agents, trading engines and design tools that measure what they did before they ask you to trust it.
Why any of this
Most software tells you it worked. Mine shows you. It measures what it built, says what it cannot, and keeps the result, even when the result is no.
Selected work
It measures what it built, and says what it cannot.
UXAI reads a live interface and the codebase behind it, plans the redesign, edits only what you approve, then opens a real browser and measures what actually rendered. This site was redesigned with it.
A council of AI agents, trading in public.
Agents debate BTC, ETH and SOL trades against the real Binance order book with paper money, on their own ledger. After 800+ published tests found nothing that beat simply holding, the main account now holds BTC by a 200-day trend rule, while an AI research engine keeps searching. The losing numbers stay public next to the winning ones.
One player, one lifelong record.
A Malaysian football identity platform: every player gets a Football ID and a QR passport that follows them from academy to club, with matches, stats and verification that no organisation can erase. Privacy is set per field, by the family.
Climate data, with automation rigour.
A platform that turns raw sustainability data into reports worth acting on, using AI and workflow automation. I am building it because emissions reporting deserves the same automation as everything else.
From dispatch to doorstep.
A laundry pickup and delivery business built from scratch: the customer apps, the logistics behind them, and the whole experience in between.
Under the hood
Every product above runs on parts I pulled out into their own libraries. Each one was born from an incident, not an opinion.
A foundation built from real production failures, each turned into a rule the code cannot break: a zero-copy transport, fail-loud config, values typed by whether they were measured or made up, and a gate that refuses to call anything proven until it is. It runs in two unrelated products.
An AI invents trading rules around the clock, a genetic breeder evolves the best of them, and a Rust engine built on THE CORE tests each one against seven years of market data. Nothing counts as a finding until THE CORE's proof gate accepts it.
Answers one question before any money moves: does this trading signal survive its own transaction costs, and exactly where does it fail?
252
hypotheses tested; the ones that failed were published
The multi-agent debate framework behind the trading council, lifted into a domain-agnostic library: provider fail-over, weighted consensus, track-record weights.
1 : 1
parity with production proven before rewiring
Carbon2Data's promo reel: a 1080p video and its synthesised score, plus a self-contained web version, all generated from code. No editing suite involved.
0
frames edited by hand
Private repositories. Walkthroughs on request.
THE CORE · measured in production
Two programs on one machine used to pass each market snapshot the slow way: 74 ms a hop. THE CORE shares the memory instead: 7.3 µs. In the time the old path moves one snapshot, the new one moves 10,108.
Before · 74 ms a hop
moving 1 snapshot…
THE CORE · 7.3 µs a hop
0 snapshots
Slowed down 40× so the old lane is visible at all. The ratio is the real one, measured in production.
● Live · the proof gate
An AI invents trading rules, a breeder evolves them, and THE CORE’s proof gate judges every one on years of market data it has never seen. This is the live record. Most ideas fail, and that is the point: nothing gets near money on hope.
Ideas judged
87,477
Proof required right now: a t-score above 5.00. The bar rises with every idea tried, so a lucky one can’t sneak through.
Why they died
Verdicts, latest run
replaying the latest run
Each verdict comes from THE CORE’s core-edge. The rules themselves stay private; only the reason they failed is shown.
The one pattern still standing
Buying the big coins right after a sharp 24-hour crash and selling a few hours later is the only idea that keeps surviving, and the gate keeps getting harder: every live rule is re-judged each run against a bar that rises with every idea tried. Its first version, which also waited for a volume spike, now falls just short and trades live on paper anyway, because only new data can settle it.
Watch it liveMethod
A vision model looked at a 240px sidebar and said 250. Twice. Then it forgot the sidebar existed, twice more. So UXAI reads geometry from the browser instead, and gets 240 every time.
240px, measured · 250px, guessed
One day a quote feed returned a fake 100.00, and the bot booked a stop against a price that never existed. Now a made-up value and a measured one are different types, and only the measured one can act.
Guarded<T>: display() for anything, act() only for the truth
Seven years of market data, twelve pre-registered gates, two passed. That result stays on the record, not in a drawer. A test that can only report success is not a measurement.
2 of 12 gates · kept on the record
Toolkit
Terminal
For the people who read source before they read copy. There are a few things in here that are not on the page.
daniel.sh: an actual shell into this site.
Try "help", "uxai", "projects" or "secret".
Contact
AI systems, automation, or a product that has to be right rather than just plausible. I answer every message myself.