Co-founder and CTO of AirShelf, an Antler-backed startup in Singapore. AirShelf turns sellers’ product knowledge and sales rules into sourced answers and offers that buyers and their AI agents can act on.
Before that I was co-founder and CTO of 12Go, APAC’s leading ground transportation OTA. I joined in 2014, and over a decade we scaled it to 50,000 bookings a day with a 45-person remote team.
Projects: notanemulator.com (DOS programs translated to C, not emulated) · fnkey · feeder · ai-9p
Karpathy’s autoresearch and DeepMind’s AlphaEvolve improve things with one simple loop: propose a change, score it, keep it only if it wins. I ran the same loop three ways: on trading strategies with a backtest as the scorer, on a spec and a strategy with LLM judges as the scorer, and by hand through 56 LLM review panels. The backtest loop inflated its own results 3 to 4 times. The LLM-judged loops stalled and learned to flatter the judges. Here is what each failure teaches, and how to build a loop that works.
South Korea put light green licence plates on expensive company cars in 2024 to stop owners from driving supercars on business tax deductions. The plates cut registrations by a third in the first year and turned into a badge of wealth in the second. The tax office now audits from the list they created.
Wise is a money-transfer network that added an account and a card. Revolut is a bank app that added transfers, and their latest annual numbers show why neither company simply copies the other.
Old DOS software usually survives inside an emulator. With vc-linux, I translated Volkov Commander and other classic DOS programs from 8086 instructions into C ahead of time, checked every instruction against a CPU emulator, and compiled the result to native Linux and WebAssembly.
The usual advice for reading protected Mac data over SSH is to give sshd Full Disk Access. That hands every SSH login your Mail, Messages and Safari history. A 154-line launcher gives the grant to one fixed program instead.
For months I have sent every Russian translation through Gemini, because Claude’s Russian reads like American English. I assumed Google’s search index was the reason. The data says Gemini really is ahead, but the lead is one year old, and it is not the index or the tokenizer.
Summing RSS says 26 Claude Code sessions eat 50GB. Measuring PSS/USS says 32.5GB — and the entire swing is the TypeScript language server, not Claude.
Claude Code’s desktop CI panel shells out to local gh in the session’s working directory — a remote path that doesn’t exist on my Mac. I fixed the premise by making the path exist: first with a FUSE-T mount that pinned a CPU at 100%, then with a git-worktree mirror that doesn’t. Plus the open-source tool that does it for you.
I asked my AI agent to plan a family escape from Chiang Mai’s burning season. It reverse-engineered a proprietary air quality encoding, scraped flights, checked fuel costs, and compared 15 cities across 10 years of data. This is what Web 4.0 actually looks like: open databases + agents that query them.
We have UX. We have DX. But AI agents are now the primary users of developer tools, and nobody’s designing for them. Here are 10 principles from watching real agent behavior through usage telemetry.