Five work examples built with Codex, each shipped as a single HTML file. No install, no server — open in any browser.
Live demos → cctail.github.io/codex-portfolio
A fully playable browser visual novel — zero dependencies, zero assets. All pixel art is generated procedurally via Canvas API (buildings, hills, characters, seasonal particles); all audio via Web Audio API (chord voicing, envelope shaping, per-heroine music themes). Three complete character routes, affection meters, branching endings, save/load, dialogue backlog, auto/skip modes.
Built through ~20 rounds of iterative Codex-assisted development. I directed Codex like a junior engineer: write the spec, review the output, catch the regressions, iterate until it ships.
A domain-expert system encoding 1,500 years of Chinese metaphysical logic in client-side JS — no backend, no dependencies, single HTML file. 2,000+ lines of pure vanilla JS implementing a real astronomical algorithm (S_TERM_INFO) for solar term calculation across 200 years, a full 60-cycle Jiazi calendar, weighted hidden stem scoring, and polarity-aware Ten God logic from scratch. Seven analysis modules, detection of 合冲刑 interactions and five 神煞 types, full Chinese/English i18n toggle, and a built-in validation panel running known-correct test cases on page load. Every output computed from first principles — no oracle, no lookup table.
An interactive dashboard defining and instrumenting developer productivity for an AI coding tool. Covers North Star metrics (session productivity, coding capacity utilization, net time saved), guardrail metrics (latency, error rate, regression), and drill-downs by language, repository size, and task type. Mirrors the measurement framework I built for Meta’s internal AI coding assistant.
A role-tailored resume built as a single HTML file — responsive two-column layout, warm paper aesthetic, clean print stylesheet that exports correctly out of the box. Content structured around LLM evaluation, developer tools measurement, and causal ML. The same self-serve clarity I’d apply to an executive dashboard, applied to a resume.
A product opportunity brief on where Codex’s next capability jump lives: eliminating the translation tax for performance marketing DS teams. Ads measurement data scientists spend 30–40% of their time translating analytical intent into working code (geo holdouts, CUPED experiments, incrementality regressions). Codex can close that gap with domain depth. Includes a concrete success metric, a proposed eval dataset design based on the LM-as-a-judge framework I built at Meta, and open questions for Day 1.
Ning Ai · cctail435@gmail.com · linkedin.com/in/ning-ai-62249a109