Josh Angel

Intelligence develops by meeting the world, encountering its limits, and finding another way.

AI researcher and engineer in Auckland, New Zealand. I research, build and run models, agents, software and DevOps infrastructure, learning from experiments and the constraints they reveal. More about me

Read the research rivet-1b paper

Explore my work

Selected work

All models & agents

parameters: 1.09B
pretraining: ~130B tokens
architecture: Qwen3-family
tokenizer: 32k SentencePiece
hardware: 2× RTX 5090 + 2× H100
inherited weights: none

Model · 2026 Paper + weights

rivet-1b, a 1.09B model trained end to end

Tokenizer, corpus, pretraining, post-training, evaluation and release. No weights, embeddings or tokenizer state inherited from another model.

SentencePiece · Qwen3 architecture · Hugging Face

Read the build note

Agent · 2026 In daily use

Ledger, agentic memory for my working life

Typed events captured upstream with source IDs and explicit human gates. The first version scraped meaning downstream and produced 66 to-dos, five of them real.

Private · macOS surface · typed event store

Read the write-up

Research & Writing

All 13 articles