CIAP Agent Assist
A real-time AI assistant that listens to live support calls, understands what the customer needs and briefs the human agent before they have to ask. Built to run entirely on-premise, with no paid cloud AI.
What it does
As a call unfolds, the assistant follows the conversation turn by turn, including interruptions, corrections and changes of topic. It recognises what the customer is asking for, gathers the details it needs, looks up the answer in the company's own systems and gives the agent a short, plain-English briefing, so the customer gets a faster and more accurate answer.
How it works
A staged pipeline that always reaches for the fastest, most reliable method first and brings in a language model only when the conversation calls for it.
- Understands the requestSemantic models recognise what the customer wants, even when it is phrased in ways the system has never seen.
- Gathers the detailsKey information is captured as the customer speaks, and missing details are tracked across the call.
- Follows the conversationA conversation state engine keeps context through interruptions, corrections and multiple requests at once.
- Escalates intelligentlyA local language model steps in only for the genuinely hard cases, with safe fallbacks so a call never breaks.
- Connects to business systemsBuilt on the Model Context Protocol, so new CRMs and tools plug in through configuration rather than new code.
- Answers with evidenceEvery briefing is checked against the retrieved data before the agent sees it.
What I built
- The end-to-end architecture and the production API, designed for many clients and many simultaneous calls.
- A pluggable integration layer with secure, standards-based authentication for third-party systems.
- Relevance-aware retrieval that keeps briefings accurate even when the underlying records are large.
- Live configuration, so behaviour can be tuned per deployment without downtime.
- Demo and replay tooling to show the system on realistic calls.
Quality
The system was developed test-first, with an extensive automated test suite, evaluation on deliberately hard questions and load testing at high concurrency. Relevance-aware retrieval improved accuracy on the hardest questions several-fold over the original approach.
Approach
Deterministic methods wherever they are cheaper and more reliable than a language model, behaviour driven by configuration, and every design trade-off measured rather than assumed.