Pomona: open-source AI for offline greenhouse decision-making

Pomona addresses a persistent tension in AI-assisted agriculture: how to deploy intelligent decision-support at the edge—where connectivity is unreliable and latency matters—while maintaining human oversight and deterministic safety guarantees. The platform's architecture (sensor → MQTT → reasoning → safety gate → dashboard) reflects a deliberate choice to keep the AI in an advisory role. This is not merely a usability preference; it is a governance requirement for any system controlling irrigation, nutrient delivery, or climate in a greenhouse where a wrong decision can destroy a season's crop.

The project's model catalog reveals the iterative nature of fine-tuning for agriculture. The tomato and irrigation reasoners are already published on Hugging Face as LoRA weights (lightweight adaptations of a base model), with version numbers and detailed regression tracking (e.g., actuator command gate v0.1.1 and v0.1.2 both flagged as regressions). This transparency—publishing failed versions alongside successes—is uncommon in commercial ML and signals a commitment to accountability. The deterministic actuator command gate, still in research preview, is the lynchpin: it must validate every LLM suggestion before any hardware action occurs.

The project is midway through its 11-phase roadmap, with phases 0 and 1 complete (setup and local MVP) and phase 2 (dashboard and production hardening) the immediate focus. Near-term integration chains sensor quality → tomato risk → safety triage → deterministic gate, suggesting the platform will gradually absorb more specialized reasoners. The Apache-2.0 license and open contribution model (including calls for help on dashboard, tests, and crop templates) indicate the authors expect community-driven improvement, particularly for crop-specific models.