Agronaut
Summary
Agronaut (github.com/Rekin226/Agronaut) is an open-source "personal agronomy agent" whose first deep domain is aquaponics. Unlike a generic chatbot that retrieves what a paper said, Agronaut computes answers for a specific system: a deterministic, fully tested engineering core (aqua_model) applies published coefficients (FAO 589, UVI/Rakocy and other cited literature) to feeding-rate-ratio sizing, nitrogen and water mass balances, and a bounded fish × crop-mix optimizer. An LLM layer only gathers facts, routes to tools, and explains results in plain language — and the Design Calculator and Optimize Ratio modes run with no LLM at all. The model covers five fish (tilapia, clarias, channel catfish, trout, common carp) and 30+ crops. Deployment targets low-connectivity field use: pluggable backends include local Ollama, free hosted open models, or a self-hosted OpenAI-compatible server, with Telegram and WhatsApp channels plus a Streamlit web app and Docker one-command startup. Built by a hands-on aquaponics operator, it learns from user-reported outcomes within literature-bounded calibration ranges. License: MIT. Actively maintained (commits through September 2026), with a stated ambition toward Digital Public Good platform independence.
Guiding Principles
- Open Source — MIT-licensed code; every coefficient ships with value, range, unit, and source so the math is auditable without trusting the LLM.
- STEM Integration — turns aquaponics design into applied mass-balance engineering, making the underlying science visible rather than hidden.
- Food Sovereignty — an offline-capable design tool (Ollama/local models) that works where connectivity and commercial agronomy services do not.
- Transferable Skills — users learn systems sizing, water budgets, and evidence appraisal (each output includes an explicit "what's NOT modeled" list).
Learning Outcomes & Transferable Skills
Learners practicing with Agronaut build Systems Thinking (feed → fish → nutrients → plants mass balances), Water Chemistry intuition (nitrogen pathways, feeding-rate ratios), Data Literacy (calibration vs. validation, cited vs. calibrated coefficients), and Food Production planning (bill of materials, operating envelopes, maintenance checklists). For Coding-oriented users, the pure-Python, fully tested aqua_model core is a readable example of validation-gated engineering software, and the repo includes a portable agentskills.io skill (skills/aquaponics-engineer/) demonstrating CLI tool design.
How to Replicate or Build On It
- Clone
https://github.com/Rekin226/Agronautand eitherdocker compose up web(Streamlit at localhost:8501) or create a venv,pip install -r requirement.txt, thenstreamlit run app.py. - Read
aqua_model/coefficients.pyfirst — the cited data layer is the project's trust core. - Run
python3 -m pytestto see the deterministic test suite (no model server needed). - Add chat by setting
LLM_PROVIDER=ollamafor fully local operation, or configure Telegram/WhatsApp per the README. - Extend the species/crop seed databases (
species.py,crops.py) with locally relevant, cited coefficients.
Considerations & Constraints
- Cost: Software is free (MIT). Optional hosted LLM tiers (NVIDIA, Hugging Face) have free tiers; fully local operation costs nothing beyond hardware.
- Skill Level: Intermediate — Docker or Python virtual environments; Telegram/WhatsApp bot setup requires tokens and (for WhatsApp) a Meta Business account.
- Space/Climate: Software only; the model targets systems from backyard to small commercial scale and handles tropical-to-temperate species.
- Tools Required: A machine able to run Python/Docker; a GPU or decent CPU if running local LLMs; internet for hosted providers.
- Ongoing Maintenance: The deterministic core needs none; calibration requires real measured outcomes from a running system. A live aquaponics system itself needs daily feeding/monitoring.
- Access Barriers: English-only documentation; coefficients are literature-seeded defaults that must be calibrated against a real system before trusting designs; some referenced sources are behind publisher paywalls.
Prerequisites
Basic Python/Docker familiarity; no vault-internal prerequisites.
