AgarAmicus-Lite

Inspiration

Re-entry of field data every session wastes time • Generic recommenders ignore soil/climate and constraints • No durable memory → poor adoption and low trust

What it does

• Top-3 crop recommendations from local conditions (NPK, pH, rain, temp, humidity, soil, irrigation) • Durable field profiles & preferences (organic/budget/irrigation) with Forget controls • MemMachine-ready memory adapter

How I built it

Frontend: Streamlit Model: RandomForest (macro-F1) Data: Synthetic (14k) Memory: JSON adapter with upsert/get/list/forget Files: data_gen_agri.py, train_agri.py, agri_app.py, mem.py

Challenges I ran into

Macro-F1 on hold-out; class report in app Educational MVP; not agronomic advice Synthetic only; no PII Requires local guidelines & soil tests in production

Accomplishments that I'm proud of

Profile memory: organic/budget/irrigation by field_id Episodic: last recommendation + last conditions Controls: list & forget Drop-in MemMachine later

What I learned

MemMachine integration (TTL, audit) Few-shot fertilizer reasoning Batch scoring & portfolio view Map layer & localized varieties

What's next for AgarAmicus-Lite

want to increase Data: Synthetic (14k)

Built With

• agri-app.py
• frontend:-streamlit-model:-randomforest-(macro-f1)-data:-synthetic-(14k)-memory:-json-adapter-with-upsert/get/list/forget-files:-data-gen-agri.py
• train-agri.py