Skip to content

Possible Pomona repositories

Where each piece of the ecosystem lives. Split-first layout (recommended).

Required repos

1. okyanu/pomona — Platform

This folder. Everything needed to run the edge AI stack locally.

services/          core, dashboard, model-router, safety-checker, automation-engine
examples/          simulators
devices/           ESP32 firmware (future)
templates/         greenhouse-tomato deployment template
models/registry/   YAML metadata only — points to Hugging Face
docs/              architecture, roadmap, publishing
docker-compose.yml

Who clones it: Platform developers, users running Docker, contributors to dashboard/automation.

Does NOT contain: Training code, checkpoints, datasets, model weights.

2. pomona-agronomist-llm — ML training (not yet published)

Extracted from models/agronomist-llm/ via ./scripts/split/extract-ml-repo.sh.

src/               digital twin pipeline
finetune/          LoRA training
deploy/            Gradio app + upload_to_hub.py
checkpoints/       local weights (gitignored)
data/              training data (gitignored)

Who clones it: You, when finetuning or updating the agronomist model.

Links to: Platform via HF model ID; platform does not need this repo at runtime.

3. Hugging Face — published models and datasets

Repo Type Publish script
Okyanus/ai-pomona-agronomist-gemma4 Advisor LoRA ML repo ./scripts/publish/huggingface.sh
Okyanus/pomona-tomato-risk-reasoner-v0.1.7-lora Tomato reasoner LoRA Published — platform ./scripts/huggingface/publish_tomato_reasoner_to_hf.sh for new versions only
Okyanus/pomona-water-irrigation-risk-reasoner-v0.1.8-lora Water/irrigation reasoner LoRA Published release candidate — ./scripts/huggingface/publish_water_irrigation_reasoner_to_hf.sh
Okyanus/pomona-actuator-command-gate-reasoner-v0.1-lora Actuator-gate LoRA Published research preview — deterministic checker remains final
Okyanus/greenhouse-sensor-data Published greenhouse sensor dataset ./scripts/huggingface/publish_dataset_to_hf.sh for future dataset releases

Platform GitHub repo keeps code, schemas, and registry YAML — not .safetensors or teacher datasets.

Future repos (add when needed)

Repo When to create Contains
pomona-tomato-reasoner If tomato training outgrows private/colab/ Training + eval for tomato reasoner
pomona-docs Only if docs outgrow platform repo Public site, tutorials
pomona-devices Only if firmware explodes (many boards) ESP32, RP2040 firmware monorepo

Do not create yet: Separate repos for core, dashboard, model-router — keep those in okyanu/pomona.

Visual map

                    ┌─────────────────────────┐
                    │  GitHub: pomona         │
                    │  (platform)             │
                    │  docker compose up      │
                    └───────────┬─────────────┘
                                │ reads registry YAML
                                ▼
                    ┌─────────────────────────┐
                    │  HF: ai-pomona-         │
                    │  agronomist-gemma4      │
                    │  (weights)              │
                    └───────────▲─────────────┘
                                │ upload
                    ┌───────────┴─────────────┐
                    │  GitHub: pomona-        │
                    │  agronomist-llm         │
                    │  (train + deploy)       │
                    └─────────────────────────┘

Split command

# From pomona/ root — moves ML code to ../pomona-agronomist-llm
./scripts/split/extract-ml-repo.sh

Then publish each repo independently. Maintainer publish checklists are local only (not in this public repo).

Summary table

# Name Platform Purpose
1 pomona GitHub Run the stack
2 pomona-agronomist-llm GitHub Train advisor LLM
3 ai-pomona-agronomist-gemma4 Hugging Face Download weights
4+ Future reasoners GitHub + HF each When Phase 3+ ships

Total GitHub repos today: 2. Not a monorepo.