Service · AI Assistants

AI that answers from what's actually true.

A copilot wired into your systems that can do things safely, or a chatbot that only speaks from your vetted knowledge — never a raw model bolt-on.

The problem with a raw chatbot

An LLM will confidently make things up. Ours doesn't get the chance.

Silent Sentinel runs a grounded sales chatbot on exactly this pattern — a closed-world knowledge base, explicit governance rules, and a 92-test harness, twenty-two of them deliberately adversarial, that a change has to clear before anyone approves it going live.

How a governed answer is built
01
Closed-world knowledge base

A curated, versioned set of source documents — nothing outside it is admissible as an answer.

02
Governance rules

Explicit ALWAYS / NEVER rules the assistant must obey on top of the knowledge base.

03
Adversarial test rig

A standing regression harness that attacks the live assistant — smoke, adversarial, and chained scenarios — before any KB or rule change ships.

04
Governed answer

Only what survives all three gates reaches the person asking.

Live Gemini GEM92-test harness, run before anything is approvedTen curated source documents — nothing else is admissible

Want a chatbot that can't wander off your knowledge base? Half an hour to see how it's built.

Start a conversation
Copilots that act, not just answer

Assistants that can reach real systems, built to the risk they carry.

Olivers — “the Vase”
Built

A back-office assistant that can act on the shop's own data — reads orders and stock, and drafts customer replies — behind a confirm-token gate before anything destructive commits.

Silent Sentinel — camera assistant
Working

An assistant embedded in the camera console, driving a catalogue of around 50 tools against a real camera — not a chat window bolted on beside the controls.

Component · Voice

Assistants can take voice as input, not just typing.

Silent Sentinel's camera assistant accepts voice alongside text, reaching the same catalogue of tools either way. Full voice-interface work is its own service.

See voice & hardware work →

Tell MAST which piece of work consumes your week. Half an hour is enough to see whether an agentic system is worth building.

Start a conversation
MAST
consulting
MAST Consulting — AI Systems & Agents.
matt@mastconsult.co.uk · mastconsult.co.uk