Software engineering at the University of Canterbury. Generally interested in high-performance computing, distributed systems and AI infra.
ARRO - Co-founder & CEO.
Building software for New Zealand's building-consent process. We analyse architectural drawings sets to preemptively raise likely RFIs before architects lodge for consent. Led a team of 5 from hackathon win to revenue in 5 months.
Halter - Incoming SWE intern (summer 2026–27).
Malaghan Institute of Medical Research (2025–26). Worked on a single-cell RNA-seq analysis platform, first as an intern and then part-time.
- Built a distributed runtime for pipeline steps on containerized GPU workers, with 5–20× speedups on datasets above ~50k cells.
- Developed DAG-based workflow execution with parallel scheduling, dependency tracking, and checkpoint rewind/resume.
- Content-addressed checkpointing reduced per-node storage from ~1.5 GB to ~27 MB (~98%). Also worked on a tool-using RAG pipeline and a sandboxed prototype for generating new analysis methods from notebooks or natural-language specs.
AFX-1000 - Rebuilding the Pioneer RMX-1000 effects unit in C++20.
- Real-time DSP engine with multiband isolation, beat-synced effects, sample playback, and 10 Scene FX.
- Native macOS app, AU/VST3 plugins, and direct audio capture from rekordbox. The audio callback runs without heap allocations or locks.
Also working on a standalone PCB version.
DAMASCUS - LLM inference engine for Apple Silicon, written in C++20 - damascus-llm.dev
- Implemented GGUF loading, byte-level BPE, Q4_K_M/Q6_K dequantization, grouped-query attention, KV-cache execution, and an end-to-end transformer forward pass.
- Checked the CPU path against a same-weights Hugging Face reference, with matching top-1 tokens and near-identical layer outputs.
Currently researching Metal execution, AMX-accelerated prefill, and CPU/GPU scheduling on M1.
Liquorfy - NZ-wide price comparison for alcohol - liquorfy.co.nz
- Built ingestion and product-matching pipelines for 62,000+ products from 12 retailers, deduplicated into ~20,000 canonical products.
- Tracks 1.6M+ prices and 6M+ price-change events across 1,400+ runs.
- PostGIS and Redis back geospatial search over 1,365 stores in 80 regions, with normalized per-100ml and per-standard-drink pricing.
- Indexed lookups measured at ~0.07 ms/row on the 1.6M-row price table.
Trying to get this one on the App Store.
-
3/3 hackathon winner · 2nd place in the Entre Enterprise Challenge.
-
Open source contributor to scverse's Scanpy
C++20 · Metal · Python · TypeScript · Swift · C# · PostgreSQL/PostGIS · Redis · FastAPI · PyTorch · Docker · React
