Principal-level contract engineering, most recently owning the AI capability of an HR-tech SaaS end to end: the agentic layer of the product — an in-app copilot with tool-calling over live customer data, autonomous document-generation pipelines, and the server-side compute services underneath — while shipping across the full stack (PHP/GraphQL, React, Go gateway, edge services).

Built an in-app AI copilot: multi-provider tool-calling agent (Anthropic + OpenAI via Vercel AI SDK) with data-query and mutation tools, an LLM-as-judge quality gate scoring every response on 8 dimensions, and SSE keep-alive for long agent streams

Designed an autonomous report pipeline — one model plans the document, Claude Agent SDK runs write it against live data, a second agent pass compiles branded slide decks via a 90+ layout PPTX compiler with chart rasterization and LaTeX PDF output, agents running in a parallel containerized fleet

Moved heavy analytics server-side (Bun + Hono behind a Go gateway): pure-function ports oracle-verified for numeric parity with the client engine, event-keyed Redis caching, validated at 20k+ entity scale

AI data enrichment: ingested a ~32-table public occupational dataset, LLM-classified it into customer data structures, and built persona-driven simulation commands generating realistic test data at scale

Shipped a standalone real-time collaborative app on Cloudflare Workers + Durable Objects — WebSocket presence, optimistic updates with server reconciliation, embedded AI assistant with client-side tool execution

Hunted down multi-user concurrent-edit data loss with a purpose-built integrity harness: truly-parallel mutation racing, event-chain fork detection, contract tests driving real client reducers against the live API

Creator & Maintainer at MobTranslate

Open-source language revitalization platform for Aboriginal Australian languages — dictionaries for dozens of languages, community voice recording, and a full speech/ML stack built with the community, not just for it.

Neural TTS pipeline: MMS-TTS (VITS) driven through a grammar-grounded orthography bridge that maps Kuku Yalanji spelling onto a same-family Pama-Nyungan donor voice (per Patz's reference grammar), per-word phonemic IPA for 2,679 dictionary words, deterministic seeded synthesis with full provenance, and a per-language model registry

ASR for low-resource speech: wav2vec2/Omnilingual CTC fine-tuning pipeline with reproducible dataset builds, serverless training runs, and benchmark harnesses

LLM translation fine-tunes with research-grade evaluation: lexical censuses, prompt ablations, dose curves, failure taxonomies, and curriculum replay datasets across Kuku Yalanji and Mi'gmaq

Native Android keyboard (IME), community recording infrastructure where real speaker audio always outranks synthesis, and sentence-corpus pipelines that turn recordings into training manifests

CARE-aligned governance: synthetic voices are explicitly flagged as scaffolds, donor voices are borrowed until a community licenses its own, and per-speaker (elder) models require explicit, revocable consent