AI Project Concept Template for Windesheim/VCH

All projects follow Agile principles with iterative development, regular sprint cycles, and continuous stakeholder feedback.

PROJECT CONCEPT

Basic Information

Project Title: ChainForge (VCH Fork)

Project Slug: chainforge

Concept Status: in-development

Category: internal-tool / research

GitHub Repository: https://github.com/Value-Chain-Hacking/ChainForge Upstream: https://github.com/ianarawjo/ChainForge

THE 3 MINUTE RULE PITCH

1. What is it?

A visual programming tool for testing and comparing AI prompts across different models without writing code.

2. How does it work?

You create prompt variations using a drag-and-drop interface, connect them to multiple AI models (OpenAI, Anthropic, Google, etc.), run them all at once, and see which prompts work best through automatic scoring and comparison charts.

3. Are you sure?

It’s an existing open-source project (chainforge.ai) with 369 commits and active community. We’re forking it to customize for supply chain research use cases.

4. Can you do it?

[TO FILL - Is this a fork we’re maintaining or contributing upstream?]

5. What’s the value?

Students and researchers can systematically test prompts instead of ad-hoc trial and error. Critical for supply chain AI projects where prompt quality directly impacts data extraction accuracy.

6. Are there any risks?

• Maintenance burden if we fork heavily
• API costs for testing multiple models
• Learning curve for students unfamiliar with visual programming

The Problem

What problem does this solve? Researchers test LLM prompts by manually copying/pasting into ChatGPT and comparing results in their heads. This doesn’t scale, isn’t reproducible, and wastes time. Need systematic “battle-testing” of prompts.

Who has this problem? VCH students and researchers using LLMs for supply chain data extraction, analysis, and research tasks.

The AI Solution

What AI/ML technique would you use? Prompt engineering, systematic evaluation, multi-model comparison. Not building AI - building tools to use AI better.

What data would you need? Access to LLM APIs (OpenAI, Anthropic, Google, etc.). Test datasets for supply chain scenarios.

Expected output/deliverable: Working ChainForge installation for VCH/Windesheim students, customized with supply chain prompt templates, documentation/tutorials for common use cases.

CURRENT STATUS (from GitHub)

Repository: https://github.com/Value-Chain-Hacking/ChainForge (Fork) Upstream Status: Active (369 commits, open beta) Created: July 25, 2024, last activity Nov 10, 2025

Tech Stack:

• TypeScript (91.1%), Python (5.2%), CSS (2.9%)
• ReactFlow (visual programming)
• Flask (backend)
• Docker support

Key Features:

• Query multiple LLMs simultaneously
• Systematic prompt variation testing
• Custom scoring functions
• Chat conversation template support
• Multi-turn dialogue testing
• Export results to spreadsheet
• Visual metrics comparison

Supported Providers: OpenAI, Anthropic, Google (Gemini, PaLM2), HuggingFace, Ollama, Azure OpenAI, AlephAlpha, Amazon Bedrock

NOTES FOR COMPLETION

Key Decisions Needed:

• Are we maintaining a fork or contributing upstream?
• What supply chain-specific customizations do we need?
• How will we provide API keys to students? (cost management)
• Do we host an instance or students run locally?
• What tutorials/templates will we create for supply chain use cases?

Potential Customizations:

• Pre-built prompt templates for supply chain data extraction
• Sample datasets (ESG reports, shipping docs, etc.)
• Tutorials for common VCH research tasks
• Cost tracking dashboard for API usage
• Integration with SupplyLens or other VCH projects

[TO FILL - Timeline, deployment plan, student training materials]