AI engineer · Software verification & validation · Australia

I came to software through automotive test engineering, so I build tools the way I was trained to test: written requirements, repeatable checks and evidence someone else can inspect.

• agent-eval-harness: compares coding models and agent CLIs on the same tasks, graded by hidden pytest suites, with unbiased pass@k and JSONL evidence.
• api-vv-toolkit: turns an OpenAPI 3.x spec and written requirements into executed API tests and a requirements traceability matrix.
• codegraph: indexes Python and Java into an embedded code graph and answers callers, inheritance, dead-code and change-impact queries with file:line evidence.
• bangla-llm-eval: a small, reproducible Bangla evaluation of LLMs (translation, NLI, Bengali numerals) with raw outputs and a leaderboard.
• humming-keyboard: Humming, a private keyboard with offline voice typing on your device, including Bangla.
• code-video-kit: 34 typed Remotion components for code and data videos, with a rendered gallery.

• sonu: offline desktop voice typing for Windows, macOS and Linux; an independent fork of Handy.
• Antigravity-Claude-Code-Proxy: a local Anthropic-compatible gateway so Claude Code can run on other model providers; extends antigravity-claude-proxy.