A senior farm manager,
on call, who reads the books.
FarmOps Assistant is a credit-metered AI, grounded in your real records, trained on a curated domain corpus, and constrained by a hard safety floor. It suggests, drafts, and explains. You decide what ships.
FarmOps Assistant
onlineThree of four units are on plan; Unit B is at risk after a 28% water dip and 12-bird mortality spike this morning.
Draft: Drinker line check · Unit B
Walks lines 3–5, logs findings, escalates if unresolved by noon.
What it can do
Six capabilities, all behind the same safety floor.
Ask in plain English
"Why did Unit B lose water at 7am?" The assistant reads your logs, weights, water, weather, and past batches — and answers with citations.
Grounded in your data
RAG over your farm's own records and a curated domain knowledge base. No hallucinated farms. Every claim points to a specific entry.
Drafts, not decisions
The AI can draft a sale, an expense, a feed order, or a recommendation. Nothing is posted until you tap approve.
Autonomy you control
Suggest · draft · auto. Money and inventory stay manual. The autonomy floor is yours to set, per farm.
Learns from you
Approve a draft once, the next one looks like it. The system tunes itself to your categories, thresholds, and tone.
Visible cost, every time
Your credit balance is always visible in the sidebar. We deduct after each run so you only pay for what you used. Top-up packs start at ₦500 and never expire.
A run, end to end
From question to record, with the user in the loop.
Every assistant interaction is a small workflow with explicit human checkpoints. We don't trust a model alone with your farm. The architecture enforces it.
Cite or refuse
If a number isn't in your data, the assistant says so. No confident guessing.
Draft, never decide
The assistant can propose a record. It cannot post one. That's enforced in the data layer, not the prompt.
Cheap by default
Nova 2 Lite is fast and economical. We reach for heavier models only when the task needs it.
Reversible
Every AI action is a draft until confirmed. Rejecting a draft is one tap, with a reason.
- 1
You
Ask: "Should I worry about Unit B's water drop?"
- 2
Assistant
Reads last 3 batches, water log, weather → returns answer with cited entries
- 3
You
Approve the suggested task: walk drinker lines 3–5
- 4
System
Creates a task on Unit B, logs it to audit, updates Today view
The stack
The boring, modern default.
Amazon Bedrock with Nova 2 Lite, Vercel AI SDK on the server, keyless OIDC to AWS, pgvector for retrieval. No model in a closet. No static keys. No LLM-shaped holes in your data.
| Model | Amazon Nova 2 Lite (us-east-1) |
| SDK | Vercel AI SDK + raw InvokeModel |
| Auth | Vercel OIDC → AWS IAM (no static keys) |
| Embeddings | Nova multimodal · 1024-dim · pgvector |
| Retrieval | Hybrid: BM25 + vector + structured filters |
| Memory | Per-farm, per-user preferences & decisions |
Credit-metered
Every run shows its cost. Set caps per farm. Get notified before you cross them.
RAG, not vibes
Retrieved over your own records and a curated domain corpus. The answer cites its sources.
Hooks for the future
Pluggable embeddings, swap-in models, your data exportable. Built to evolve, not to lock in.
The assistant remembers — only what you teach it.
Per-farm preferences. Per-user tone. Decisions logged so the next draft is closer to the one you'd write. Nothing leaves your farm.
Walk a real farm in 60 seconds.
The demo loads a real Nigerian farm with live insights, feed runway, money, and a chat with our AI. You'll know if FarmOps is right for you before you sign up.