KEIRAN
MOORES

An Outlook add-in running on Azure across desktop, web, and New Outlook. It files emails and attachments into SharePoint without the user leaving the inbox — and suggests where each one should go.

The recommender ranks candidate destinations by similarity against the history of prior filing decisions, returning a confidence score for each. It sharpens as precedent accumulates.

• Chose statistical matching over a trained model deliberately — the labelled data did not exist yet.
• Instrumented every decision instead: document, destination, suggestion, accepted or rejected.
• That log builds the training set and doubles as the accuracy baseline for a later move to Azure ML.

A Next.js platform with two sides. Donors are matched to NGOs; NGOs are matched to grant-making foundations. Both paths read from one shared organisation dataset, so every organisation added serves both directions at once.

The primary path is an eight-step quiz scored by a deterministic weighted engine, returning the top five matches with a plain-language reason for each. No model involved — deliberately.

• A donor deciding where money goes needs to see the reasoning, not a number from a black box.
• A parallel conversational path runs on the Claude API with separate donor and NGO modes.
• Every response is grounded in an injected organisation directory and constrained against recommending anything outside it.
• Prompt caching keeps the grounding context off the per-turn cost; responses stream token-by-token into the embedded widget.

Visit the site ↗

Research into what changes inside an SME when AI-assisted filing arrives, and what conditions decide whether people trust the recommendations enough to accept them.

Interview data is triangulated against the acceptance-rate log from the add-in — every suggestion accepted or rejected. Trust gets measured by behaviour rather than by asking people how much they trust it.

• Filing is framed as tacit knowledge: folder conventions encode expertise held by individuals and teams, rarely written down.
• The question is whether a system can externalise that knowledge reliably enough to be trusted with it.
• The field site is the organisation the add-in was deployed into — so the log and the interviews describe the same people.