An AI consultant that helps educational software teams turn an idea or pain point into a concrete, implementation-ready A/B experiment plan for UpGrade, an experimentation platform for educational software.
You describe your learning app and something you'd like to improve, and the consultant guides you — in a single conversation — to a plan you can act on:
- Guided consulting flow — app description → page/problem → hypothesis → experiment design → preflight simulation → final report.
- Screenshots & files — attach images or PDFs to ground the discussion.
- Research grounding — pulls related papers from Semantic Scholar to help sharpen a hypothesis.
- Preflight simulation — runs a synthetic cohort against UpGrade's demo backend so you can see how assignment, enrollment, and metrics would look (no real users).
- Shareable report — a structured markdown plan, including UpGrade setup and client-integration guides, ready to hand to researchers, developers, or an AI coding tool.
See docs/spec.md for the product vision and docs/mvp.md for scope. The PELE 2026 paper is available at paper/ai-experiment-consultant-pele-2026.pdf.
- Frontend: Vite + vanilla JavaScript (client/)
- Backend: Node.js + Express, ES modules (server/)
- AI: Anthropic Claude — streaming chat with tool use
Requirements: Node.js 20+. Running the full app also requires Anthropic, Google OAuth, UpGrade service-account, and Semantic Scholar credentials; see server/.env.example.
# Clone and install (installs both the frontend and the server)
git clone https://github.com/CarnegieLearningWeb/ai-experiment-consultant.git
cd ai-experiment-consultant
npm install
cp server/.env.example server/.env # then fill in your credentials
npm run devThen open the URL Vite prints — by default http://localhost:5173/ai-consultant/.
ai-experiment-consultant/
├── client/ Vite + vanilla JS frontend
├── server/ Express backend (API under /api/v1/ai-consultant/*)
└── docs/ Product spec, architecture, setup, and reference notes