AI assistant · RAG · on-device ML · preview

Mercatus Assist

Name: Mercatus Assist
Author: Katlego Mokoto

Ask anything about Katlego's work. Retrieval and intent routing run on your device with a small embedding model; Gemini writes the answer from the sources it was given, and every claim links back to the page it came from. When it doesn't know, it says so.

What's inside

LIVEGrounded answers. Hybrid BM25 + MiniLM retrieval over the site's knowledge base, one Gemini call, citations checked before anything is shown.
ON DEVICEIntent router. Nearest-centroid + 2-NN over MiniLM vectors: recruiter, client, tech, booking. Each gets its own answer style and actions.
ON DEVICEConfidence meter. Computed from retrieval similarity, so a weak match is visible instead of hidden.
LIVERecruiter fit. Paste a job post: requirement-by-requirement match against the real CV, honest gaps, then a written summary.
LIVEExplain a project. Reads a case-study page and explains it for a recruiter, a business owner or an engineer.
ON DEVICEArchitecture lens. Draws a live data-flow diagram of the project being discussed.
ON DEVICESkill evidence. Each claimed skill links to the case study, app or dated build-log entry that shows it.
LIVEThree lengths, four languages. English, isiZulu, Afrikaans or Sesotho (translated by Gemini, marked as such).
ON DEVICEVoice. Speak a question; hear the answer in an en-ZA voice where your device has one.
PREVIEWStreaming. Answers are revealed progressively; true token streaming is a prepared server change.

Privacy

The conversation lives in this browser tab only (sessionStorage) and the Clear button wipes it. No cookies, no tracking. Your question and the matched site text go to the existing Assist API to get an answer; nothing new is stored on a server.

Go deeper

Case study Technical overview (PDF) All work Original policy-coach chat