The MINDS (Mandela Institute for Development Studies) Scholarship Programme has produced a growing network of alumni spread across the African continent and beyond, but they lacked a shared digital home. There was no central place for alumni to discover one another, surface opportunities, celebrate achievements, follow news and events, or engage new prospective scholars. Information about MINDS Africa and the scholarship was scattered across documents and web pages, making it hard for visitors to get quick, reliable answers. The organisation also needed to update all of this content themselves, without depending on a developer for every copy or image change.

I designed and built a modern, content-managed marketing and community website that acts as the network's digital hub. Every meaningful piece of content, page copy, images, regions, events, gallery, alumni ventures, modals, and even SEO metadata, is editable by non-technical staff through Sanity Studio. To solve the scattered information problem, I added a Retrieval-Augmented Generation (RAG) chatbot, an Alumni Assistant trained on MINDS' own content (CMS entries, curated knowledge documents/PDFs, and selected pages from minds-africa.org). It gives visitors instant, grounded answers about the programme, regions, events, and how to get involved, instead of forcing them to hunt through pages.

• Fully CMS-driven site

  All pages, imagery, and content blocks are editable in Sanity with sensible code fallbacks.

• AI Alumni Assistant (RAG chatbot)

  Streaming, source-grounded answers scoped strictly to MINDS topics, with suggested prompts and graceful guardrails.

• Interactive Africa scholars map

  SVG map of Africa with per-country hover tooltips showing alumni counts, driven by CMS data.

• Stripe donations

  In-app embedded checkout modal, no full-page redirect.

• Newsletter capture

  Scroll-triggered subscribe modal timed to the Stories From Our Alumni section.

• Alumni ventures & employers

  Toggleable, editable showcase of companies alumni founded vs. companies they work at.

• Regions, events, gallery, news

  Structured, filterable content sections.

• SEO & discoverability

  Per-page meta/Open Graph/Twitter tags, JSON-LD, sitemap, and robots.txt.

• Polished UI

  Glassmorphism, theme-aware navigation, and motion-based micro-interactions.

• Grounding the chatbot in MINDS-only knowledge

  Built an ingestion pipeline pulling from Sanity, editor-uploaded PDFs/knowledge docs, and selected external pages, chunk, embed, and upsert into Upstash Vector. At query time retrieve relevant chunks and constrain the model with a scoped system prompt; tuned the prompt to strip unwanted in-text citation markers.

• Provider migration mid-build

  The chatbot originally ran on OpenAI but hit quota limits. Because it was built on the AI SDK's provider-agnostic abstraction, migrated the entire chat + embeddings stack to Google Gemini with minimal changes, keeping embedding dimensions compatible with the existing vector index.

• Serverless functions in local dev

  Vite doesn't natively serve Vercel's /api routes, so the chatbot 404'd locally. Wrote a custom Vite dev plugin that intercepts /api requests, loads env vars, and executes the actual function handlers, true parity between local and production.

• Production-only outage after adding env vars

  Live chatbot failed with a generic error even though keys were set in Vercel. Diagnosed that Vercel bakes env vars in at build time, so the running deployment predated the new keys. Triggering a fresh deploy resolved it; verified by hitting the live streaming endpoint directly.

• Empowering non-technical editors

  Modelled everything in Sanity with fallbacks so staff can safely edit copy, imagery, toggles, and even the chatbot knowledge base without touching code.

Shipped a live production platform at alumni.minds-africa.org serving as the MINDS Alumni Network's central hub. Delivered a working, source-grounded AI assistant that streams answers about MINDS in real time. Handed the organisation full content autonomy, virtually every element is CMS-editable, with a self-updating chatbot knowledge base via scheduled re-indexing. Made the site search-discoverable with complete SEO metadata, sitemap, and structured data.

1. 1.Provider-agnostic abstractions pay off. Building the AI layer on the AI SDK made swapping OpenAI for Gemini a low-risk change instead of a rewrite, design around interfaces, not vendors.
2. 2.RAG quality is mostly a data + prompt problem. The hardest parts weren't the model calls but curating sources, chunking sensibly, and tightening the system prompt to keep answers grounded and on-brand.
3. 3."Works locally" needs deliberate parity. Investing in a dev plugin for serverless routes and understanding build-time vs. runtime env vars prevented an entire class of confusing production-only bugs.
4. 4.Content modelling is a product decision. Thinking through what editors need to change, and providing fallbacks, turned the site from a static build into a tool the organisation actually owns.

Live project

alumni.minds-africa.org/