CraftPilot: Building a Multi-Agent Craft Business Assistant with a Single Small Model

kamath

skamathramesh

Follow

build-small-hackathon

The Problem

I know someone who crochets, embroiders, paints, and sews the most beautiful things. Friends and family constantly tell her: "You should sell these!" But she never does. Not because the work isn't good enough — it's because the selling part is overwhelming.

Writing product descriptions? Agonizing. Picking a fair price? Impossible. Crafting Instagram captions? Exhausting for an introvert. So the crafts pile up, gifted away or tucked into drawers, while she moves on to the next project.

I built CraftPilot to fix that.

The Solution

CraftPilot is a photo-in, listing-out tool. Upload a photo of your handmade craft, and you get:

• Catalog metadata — category, materials, colors, complexity, searchable tags
• Product copy — a title, short description, full description, and 3 Instagram captions
• Fair pricing — a price range based on material cost, labor time, and market rates
• Downloadable listing — export everything as an Etsy/Instagram-ready text file
• Agent traces — full transparency into what each AI agent did

The key constraint: everything runs on a single model (MiniCPM-V 2.6, ~8B parameters) via llama.cpp. No cloud APIs. No subscriptions. No sending your craft photos to OpenAI.

How It Works: The Multi-Agent Pipeline

CraftPilot uses a 4-agent pipeline, all powered by the same model:

Photo -> [Vision Agent] -> [Cataloger Agent] -> [Copywriter Agent] -> [Pricer Agent]

1. Vision Agent — Takes the photo and produces a detailed text description of the craft item (materials, colors, techniques, style)
2. Cataloger Agent — Reads the description and outputs structured metadata as JSON (category, materials, tags, complexity)
3. Copywriter Agent — Takes the catalog data and writes warm, authentic product copy and Instagram captions
4. Pricer Agent — Considers materials cost, labor hours, complexity, and market rates to suggest a fair price range

Each agent has a specialized system prompt and outputs structured JSON via constrained generation. The pipeline streams results — you see the vision analysis appear first, then catalog data fills in, then copy and pricing.

What I Learned

One Model, Four Agents

MiniCPM-V 2.6 is a multimodal model from OpenBMB that handles both vision and text generation. Running it via llama.cpp means no GPU required, no API costs, and full privacy — your photos never leave your machine.

Small Models Need Focus

A small model is surprisingly capable when you give it:

• Clear, focused system prompts (one job per agent)
• Structured output constraints (JSON schema)
• Pre-computed math (the model can't reliably add, so the pricing template does the arithmetic)

Where it struggles: complex reasoning, nuanced pricing logic, and occasionally inconsistent JSON. Error recovery handles this gracefully — if one agent fails, you still get results from the others.

Streaming Changes Everything

The pipeline streams results as each agent completes. Users see a progress stepper and results filling in piece by piece. This turns a 30-60 second wait into an engaging experience. Without streaming, users assume the app is broken after 10 seconds.

The Stack

• Model: MiniCPM-V 2.6 (~8B) via llama-cpp-python
• UI: Gradio 6.x with custom CSS
• Orchestration: Python pipeline with streaming generators
• Hosting: Hugging Face Spaces (CPU)
• Cloud APIs: None

Try It

Try CraftPilot on Hugging Face Spaces

Built for the Build Small Hackathon 2026 — Backyard AI track. Single model, no cloud APIs, full agent transparency.

Built with love for someone who deserves to share her craft with the world.