Last updated: October 2026

I'm Johnny Kao.

This GitHub is a second Johnny: it is the more experimental, technical, and personal side of me — selected code, open-source contributions, passing curiosities, and ideas I feel like keeping around.

I spend most of my time around finance, strategy, cross-border questions, data, research, and digital systems, but I have never liked staying in just one lane.

Recently, I have been spending more time on upstream open-source engineering, especially around performance, correctness, concurrency, and low-level infrastructure.

A recurring pattern in the work is simple: find an expensive or unsafe path, establish the behavioral boundary, reduce it to the smallest defensible change, benchmark or differential-test it, and upstream it.

• NumPy #32802 — eliminated a 32-bit signed-shift undefined-behavior case; UBSan failure reproduced before and absent after the fix. Validation: 48,592 tests passed (full non-slow suite).
• SciPy #26225 — fixed incorrect iteration reporting for 4 bracketing solvers across 2 endpoint-root cases (0 iterations expected); 178 tests passed in the affected test module.
• pandas #69916 — skipped redundant NA-membership checks on 1,000,000-element numeric-conversion workloads, reducing direct benchmark time by 16.5–24.7% (not end-to-end CSV performance).
• c-blosc2 #805 — avoided unnecessary thread startup and worker wakeups; 64 B: 35.86 → 2.02 μs, 4 KiB: 38.90 → 3.63 μs in targeted benchmarks; 1,616/1,616 tests passed.
• python-blosc2 #728 — removed redundant full-block zeroing in the NumPy miniexpr gather path; ~2–15% faster in targeted workloads, with 1,596 relevant tests passed.
• Memray #1035 — improved Linux/glibc Tracker contention handling; measured runtime reductions of 13.1% at 4 threads, 24.7% at 16, and 28.0% at 256 in an allocation-heavy workload.
• urllib3 #5287 — optimized single-value HTTP header processing; ~9–13% faster request construction (0.64–1.36 μs/request saved) in tested 4–12-header workloads; 15/15 A/B comparisons faster.
• urllib3 #5304 — avoided redundant gzip output copying; saved 26.1–27.0 μs per 1 MiB of decoded output in tested single-output cases, with decoder-level gains of 14.7–39.2% across tested sizes and payloads.

Additional merged correctness work includes python-blosc2 #713 — fixed 2 NDArray read-path thread-safety bugs under free-threaded Python; focused concurrent stress repeated 3×.

Performance figures are specific to the linked PRs' test environments and workloads, not general production-wide improvements.

For a complete record of merged contributions and ongoing research, see CoreFoundry Project.

• 2026 — 📊 Deepened upstream OSS work across scientific Python, systems, networking, profiling, and compression
• 2025 — 🚀 Moved from consulting into startup finance and corporate development, supporting a multi-billion-JPY fundraising extension and board-level decisions
• 2024 — 🧭 Worked across digital transformation, AI adoption, future banking strategy, and RWA tokenization
• 2022–2023 — 🌐 Worked on institutional digital assets, wealth management, payments, and emerging financial infrastructure
• 2020 — 🧊 Code preserved in GitHub’s Arctic Code Vault
• 2019 — 🏦 Worked on virtual banking and cross-border financial infrastructure
• 2015–2018 — 💳 Built and worked around payments, Bitcoin-linked card infrastructure, fintech, SaaS, and digital-asset systems
• 2013 — 🤝 Volunteered in Tunisia

Some repositories here are serious engineering work. Some are experiments. Some exist because I wanted to understand one obscure problem properly.

That distinction is intentional.