Models & agents
What I build and run for myself, on my own hardware. Models are the capability layer; agents are the operating layer that decides what to read, what to ignore, and when a human has to stay in the loop.
rivet-1b
A 1.09B parameter reasoning language model trained entirely from scratch: custom tokenizer, custom pretraining corpus, custom post-training. No weights, embeddings or tokenizer state were inherited from any existing model.
- parameters
- 1.09B
- architecture
- Qwen3-family
- tokenizer
- 32k SentencePiece
- pretraining tokens
- ~130B
- hardware
- 2× RTX 5090 + 2× H100
- inherited weights
- none
ledger agent
My working record, brought into one interface: profile asks, sourced event history, owner-approved goals and generated leadership documents. The engine schedules workflows through a configured agent harness; macOS and web clients surface the results. The current source is a developer preview, not a public release.
- type
- agentic memory
- stores
- code, vault, engine state
- decisions
- candidate approval
- clients
- macOS + web + phone