FarmGrow

FarmGrow is an AI farming assistant — a chat workspace that answers questions on crops, soil health, pests, weather, and sustainable farm planning. Built on the Vercel AI SDK with switchable models (Claude, Gemini, GPT), it runs a structured "Agricultural Strategist" persona that returns a consistent diagnosis → yield-gap → integrated-strategy → economics → risk format. I built it full-stack: the XML-driven prompt system, the streaming chat API, a branching conversation model on MongoDB/Prisma, tool-call and token/latency telemetry, shareable sessions, and the UI.

Visit project

Client: FarmGrow
Year: 2026
Category: AI
Status: Live project

FarmGrow project snapshot — Snapshot / FarmGrow / 2026

Context

General-purpose chatbots give farmers vague, unstructured advice. FarmGrow constrains a multi-model LLM into an agricultural-consultant workflow — always reasoning through soil, water, pest pressure, and economics, and always answering in the same actionable structure — so a farmer or extension officer gets decision-ready guidance rather than prose.

Constraints

01Advice must be structured and decision-ready — a strict response schema (diagnosis, yield-gap, integrated strategy, economics, risk) the model can't skip.
02The assistant must not fabricate — grounding, citation, and missing-context-gating rules baked into the prompt files, with a pre-finalize verification loop.
03Conversations need edits and regenerations, not a flat log — messages modelled as a branching tree (parent/child, turn/role ordering).
04Serverless MongoDB cold-starts were adding ~15s per request — removed a redundant `$connect()` and tuned the connection pool.
05One chat surface, multiple model providers (Claude, Gemini, GPT) with per-session config.

Role & Team

Role: Full-stack
Team: FarmGrow
Stack: Next.js, TypeScript, Vercel AI SDK, MongoDB, Prisma
Engagement: AI farming assistant in one chat workspace.

Process

Persona as engineered files, not a string

I split the system prompt into versioned XML — identity, knowledge, task, constraints, methodology, response_schema — loaded by a prompt loader, so the agricultural persona and its output contract are maintainable and testable rather than a wall of text. Every answer is forced through the same five sections, with a verification loop that checks soil-health and IPM thresholds before finalizing.

// src/prompts/promptLoader.ts — persona assembled from versioned files
const SYSTEM = [
  "identity", "knowledge", "task",
  "constraints", "methodology", "response_schema",
].map(load).join("\n\n");
// response_schema enforces: diagnosis -> yield_gap_analysis ->
// integrated_strategy (IPM) -> economic_impact -> risk_mitigation

A chat backend that remembers structure

I built the streaming chat API on the Vercel AI SDK over a MongoDB/Prisma model with branching messages, persisted tool invocations, per-run token/latency telemetry, auto-generated titles, and shareable sessions — so a conversation is an editable tree, not a flat log.

// branching conversation tree (prisma/schema.prisma)
model Message {
  turnIndex       Int
  roleOrder       Int     @default(0)
  parentMessageId String? @db.ObjectId
  parentMessage   Message?  @relation("MessageBranch", fields: [parentMessageId], references: [id])
  children        Message[] @relation("MessageBranch")
  toolInvocations ToolInvocation[]
  @@index([sessionId, turnIndex, roleOrder])
}

Make it fast and shippable

I traced a ~15s per-request delay to a redundant Prisma $connect() on every query, removed it (Prisma connects lazily on first query) and added connection-pool config — then rounded it out with streaming markdown (code, math, mermaid), SEO/PWA metadata, and dynamic social images.

// before: forced a new connection on every request (~15s cold start)
await prisma.$connect();
return await action();

// after: Prisma auto-connects lazily on first query
return await action();

Outcome

The measurable shape of the work.

LLM providers switchable (Claude, Gemini, GPT)

Structured sections every answer follows

~0s

Per-request DB latency removed

Lessons

Prompt engineering belongs in version-controlled files with an enforced output contract, not a string literal — that's what turns a general LLM into a domain expert you can actually maintain.

FarmGrow

Persona as engineered files, not a string

// src/prompts/promptLoader.ts — persona assembled from versioned files const SYSTEM = [ "identity", "knowledge", "task", "constraints", "methodology", "response_schema", ].map(load).join("\n\n"); // response_schema enforces: diagnosis -> yield_gap_analysis -> // integrated_strategy (IPM) -> economic_impact -> risk_mitigation

// branching conversation tree (prisma/schema.prisma) model Message { turnIndex Int roleOrder Int @default(0) parentMessageId String? @db.ObjectId parentMessage Message? @relation("MessageBranch", fields: [parentMessageId], references: [id]) children Message[] @relation("MessageBranch") toolInvocations ToolInvocation[] @@index([sessionId, turnIndex, roleOrder]) }