🌱 Pomona
Open edge AI platform for agriculture — MQTT ingest, reasoning, deterministic safety checks, and dashboards for greenhouse and hydroponic systems.
Anyone can clone, use, fork, and contribute — free under Apache-2.0.
Website · GitHub · Hugging Face · Automation API · Quickstart · Contribute
🌱 Try the guarded greenhouse & hydroponic demo in your browser — no install, no login.
🎯 Try to fool Pomona — break a sensor in a pretend greenhouse and see if Pomona catches it. No install, no login.
Early MVP. Simulated sensors → SQLite-backed API → guarded reasoner pipeline → monitoring dashboard with manual suggestion decisions. No actuator execution. Docker remains the recommended deployment path; local validation also works without Docker. What to expect →
🚀 Quickstart (Docker)
git clone https://github.com/okyanu/pomona.git
cd pomona
cp .env.example .env
./scripts/setup.sh # optional first-time check
./scripts/up.sh
./scripts/sim.sh # new terminal
curl http://localhost:8080/health
python3 examples/tomato_risk_quickstart.py # real sensor JSON -> risk labels
Requirements: Docker Desktop
Optional: Python 3.9+ for simulator/tests
Full guide: docs/INSTALL.md
Target: one-command full stack (UI + DB in Docker) — docs/PLUG_AND_PLAY.md
Local validation details: docs/LOCAL_VALIDATION.md
Local validation without Docker
When Docker Desktop is unavailable, run the four-service smoke test with the
local Python environments:
./scripts/run_local_validation.sh
# or
make local-validation
Run the complete local service test suite with:
make test-local
Run tests and the full local service smoke test together with:
make local-check
It uses temporary SQLite and audit state, verifies both high-risk and routine
paths, checks the dedicated Safety Checker, and removes temporary processes and
files when it finishes.
🏗 Architecture
Sensor / Simulator → MQTT → pomona-core → DB → model-router
→ reasoner / advisor LLM → safety-checker → automation-engine → dashboard
Hard rule: the LLM advises — it never directly controls actuators. Deterministic rules make every decision, and there is no actuator execution path: automation only raises suggestions for a person to approve or reject. The rules are tested but not field-calibrated.
More detail: docs/architecture.md
🤝 Open source & contribute
Good first contributions: tests, docs, simulators, crop templates, dashboard polish.
Fork → branch → make test → open PR.
🔗 Related open repos
Model catalog and release status: docs/MODEL_CATALOG.md
🧠 Model status
Pomona keeps platform code, routing, deterministic safety logic, and metadata in GitHub. Model weights and published datasets live on Hugging Face.
| Model | Current status | Use now? |
|---|
Tomato risk reasoner v0.1.7 | Published on Hugging Face, experimental: learned a training shortcut and answers [] on 146 of 153 varied benchmark cases (details) | ⚠️ Research only; the guarded route uses deterministic rules |
Water/irrigation reasoner v0.1.8 | Published release candidate; evaluated only on synthetic, rule-derived cases | ⚠️ Advisory only, behind deterministic validation |
Sensor quality reasoner v0.1.1-boundary | Local candidate, not published | ⚠️ Owner's local integration tests only; not available publicly |
Safety triage reasoner v0.1 | Local candidate, not published | ⚠️ Owner's local integration tests only; not available publicly |
Actuator command gate v0.1 | Published research preview, below standalone release gate | ⚠️ Advisory only; deterministic checker required |
Actuator command gate v0.1.1-hardcases | Regression in local eval | ❌ Do not use |
Actuator command gate v0.1.2-correction | Regression on a held-out (mostly generated) eval | ❌ Do not use |
Nutrient / pH-EC reasoner v0.1.1-correction | LoRA, GGUF, and MLX published; evaluated only on synthetic, rule-generated cases | ⚠️ Only through the deterministic guarded route; the rules make the decision |
Near-term integration chain:
sensor-quality
-> tomato-risk
-> safety-triage
-> deterministic actuator-command gate
-> dashboard / human approval
Future model families are tracked in docs/ROADMAP.md and docs/SMALL_MODEL_FACTORY.md.
📊 Project phases
Platform v0.1.0-alpha.1 · 11 phases · 2 completed · 8 active/partial · Phase 2 is the primary product focus
| Phase | Name | Status |
|---|
| 0 | Project setup & open source | ✅ |
| 1 | Local MVP (Docker, core, simulator) | ✅ |
| 2 | Dashboard + database | ⏳ partial — local SQLite, dashboard, audit, Docker, simulation, and backup validation complete; optional sensor-ingestion auth added; broader hardening remains |
| 3 | Tomato reasoner | ⏳ partial — published adapter + guarded API route |
| 4 | Safety checker | ⏳ partial — deterministic tomato/actuator gates |
| 5 | LLM advisor | ⏳ partial — published adapter + router contract |
| 6 | Automation engine | ⏳ partial — YAML rules, dashboard evaluate/approve/reject, bounded SQLite history in local Compose; public API demo remains ephemeral |
| 7 | Public demo | ⏳ partial — static guarded demo and Try to fool Pomona game live |
| 8 | ESP32 devices | ⬜ — logger firmware compiles; not yet bench-tested, no field data |
| 9 | Model registry | ⏳ partial — public YAML registry + lifecycle metadata |
| 10 | Fine-tune pipeline | ⏳ partial — builders, validators, Colab, clean eval |
Details & update rules: docs/PHASES.md · PROJECT_STATUS.md · VERSIONING.md
📄 License
Copyright 2026 Pomona Contributors. Licensed under Apache-2.0.
