Farmer Parcel Assistant

Problem

An AgriTech team needed farmers to ask about parcels in plain language, without a form-heavy admin tool, while keeping answers grounded in real parcel data.

Tech stack

JavaScriptNode.jsDockerReactExpress.jsPostgreSQLTailwindCSS

What I built

Built a conversational backend for parcel lookup, phone linking, and message validation.
Handled common intents with validation and regex, and used LLM context only for ambiguous requests.
Deployed the stack with Docker Compose on AWS EC2.
Added a lightweight frontend to visualize message flow while the API handled the core work.

Decisions

Key technical choices, alternatives considered, and trade-offs.

Regex first, LLM second

Context

Most questions repeated a small set of parcel lookups. Routing every message through an LLM raised cost and made incorrect answers harder to control.

Alternatives

LLM on every message: more natural language coverage, less predictability and higher cost.
Forms only: safer for the backend, poor fit for farmers who needed chat.

Decision

Classify common intents with validation and regex, and pass ambiguous requests to an LLM for context.

Trade-off

Less open-ended coverage early on, with more predictable answers on the common path.

Docker Compose on a single EC2 host

Context

The assignment needed a reproducible demo under a short delivery window, not a multi-service platform.

Alternatives

Managed containers from day one: cleaner operations later, more setup than the brief allowed.
Bare processes on the host: fast to start, harder to rebuild cleanly.

Decision

Ship the API and dependencies as one Docker Compose stack on EC2.

Trade-off

Simple to deploy and demo, with less horizontal scale than a managed container platform.