llamaclaw — terminal-first scientific computing¶

A polyglot stack for reproducible causal inference, epidemiology, and machine learning. Zero cloud dependency. Runs on whatever hardware you have — laptop, server, or a 16 GB Raspberry Pi.

brew install llamaclaw/tap/esml
# or:
pip install git+https://github.com/llamaclaw/esml.git

By the numbers: 8,872 statistical functions, 25,279 tests passing, 41 built-in Canadian public-health datasets, 15 repos in the umbrella.

Where to start¶

I have a dataset → give me causal estimates

A 10-minute walkthrough from pip install to an ATE with bootstrap CIs.

Your first analysis — 10 minutes

I want to deploy on a Pi

Hardening, systemd units, Ollama + Perseus on zeus.local.

Raspberry Pi 5 deployment

I want to train an LLM + compress it

End-to-end: autoresearch pretraining → turboquant PTQ → Ollama serve.

End-to-end: pretrain → compress → serve

I want to understand the architecture

How all 13 repos fit together; cross-repo dependency graph.

Architecture

60-second quickstart¶

from esml.data import load_dataset
from esml.fn import ate

df = load_dataset("ccs_2022_2023")   # built-in Canadian Cannabis Survey
result = ate(
    df,
    treatment="cannabis_past_year_use",
    outcome="mental_health_good",
    covariates=["age_group", "province", "sex", "education"],
    method="ipw",
    bootstrap=True,
)
print(result.summary())
# Average Treatment Effect (IPW)
# estimate: -0.0432, 95% CI: (-0.0799, -0.0065), n: 17,842

The ecosystem¶

Repo	Role	Primary language	Package
llamaclaw/esml	Core package — 8,872 statistical functions, causal inference, DML, 41 Canadian public-health datasets	Python + R + C	pip, Docker
llamaclaw/turboquant	KV-cache compression (ICLR 2026)	Python + C	pip (git VCS)
llamaclaw/perseus	LLM relay — Ollama + FreeAPI + Gemini + OpenAI chain	Python	pip + ollama.com/llamaclaw/perseus
llamaclaw/autoresearch	Autonomous LLM pretraining (MPS main; CUDA on `upstream`)	Python	—
llamaclaw/tide	Terminal IDE	Go	ghcr.io Docker
llamaclaw/kronos	ARM64 terminal multiplexer (Pi-native)	Rust	ghcr.io Docker
llamaclaw/luci	Pi-side AI agent orchestration	Shell	—
llamaclaw/yoda	Claude co-architect’s persona + research journal	Markdown	—
llamaclaw/deploy	macOS launchd + Pi systemd configs	Shell + unit files	—
llamaclaw/installer	Homebrew tap + `curl \| sh`	Shell + Ruby	Homebrew
llamaclaw/llamaclaw	Umbrella — submodules + release scripts	Markdown + Shell	—
llamaclaw/intersphinx	Ecosystem documentation (this site)	Sphinx + Markdown	GitHub Pages

Plus :ollama:`` (model registry) for Perseus LLM weights.

Design principles¶

Zero cloud dependency. Everything runs locally or on a single Pi.
Dataset-agnostic. Every statistical function takes column names as keyword arguments. BYOD or use the 41 bundled datasets.
Dual-language. Python and R agree by construction.
Terminal-first. The TUI (tide/kronos) is primary; no web app.
ARM + x86 equals. The Pi is a numerical audit rig, not a deploy target.
Temperature 0.1. Scientific computing doesn’t need randomness.

License¶

GPL-3.0-or-later. Every sub-repo uses the same license.