Pomona architecture

Local-first edge platform for greenhouse and hydroponic automation — inspired by Home Assistant, built for agriculture.

Data flow

Devices / Simulator
       ↓ MQTT
  Pomona Core          ← ingest + API (✅ today)
       ↓
  Model Router         ← deterministic specialists + optional guarded models
       ↓
  Safety Checker       ← deterministic blocking and human-review flags
       ↓
  Automation Engine    ← suggestions + recorded decisions, no execution
       ↓
  Dashboard            ← monitoring, forecast previews, manual decisions

Rule: The LLM advises — it never directly controls actuators.

Services

Service	Role	Status
core	Sensor ingest, storage, REST API	✅ MVP
model-router	Route tasks to models / rules	✅ MVP
dashboard	Monitoring, previews, manual suggestion decisions	⏳ Implemented locally; Phase 2 partial
safety-checker	Block unsafe recommendations	⏳ Deterministic checks implemented; Phase 4 partial
automation-engine	YAML rules → suggestions	⏳ Phase 6, partial
digital-twin	Illustrative forecast scenarios, not hardware commands	⏳ Local preview

Core uses SQLite for sensor events. Automation uses a separate SQLite file on a named volume in Docker Compose, with bounded history retention. Neither the model nor a recorded approval has a hardware execution path. Local operation does not provide authenticated reviewer identity.

First MVP (running now)

greenhouse_tomato simulator → MQTT → core → model-router (stub)

Run: ./scripts/up.sh then ./scripts/sim.sh

More detail

Maintainers: full architecture notes in private/planning/ (local, not in Git).

Public progress: PROJECT_STATUS.md · PHASES.md