Realtime live-chat agent console with an AI reply assistant

Problem

Live-chat support agents often handle several conversations at the same time. Without help, a visitor who has been waiting too long is easy to miss, the agent has to piece together who the visitor is from separate places, and every reply is typed from scratch. There is also a quieter failure: when an agent's laptop goes to sleep, its realtime connection drops without warning, and messages that arrived in the meantime can be missing from the screen when it wakes, so a conversation looks answered when it is not.

Goal

One realtime workspace where an agent can run many chats at once, see the context for each visitor, be reminded before anyone waits too long, get AI reply suggestions they stay in control of, and never lose a message to a sleeping laptop.

My role

Worked on the agent console as part of a small development team.

Solution

A realtime agent console built around four ideas.

Many conversations, one view. Each open chat sits in the workspace with the visitor's context beside it, so switching between conversations does not mean switching tools.

Escalating wait reminders. The longer a visitor waits for a reply, the more insistent the reminder becomes.

An AI reply assistant with a human in the loop. The model suggests a reply; the agent decides whether to use it, edits it if needed, and rates it.

Wake recovery. The console notices when the computer has been asleep, fetches the messages it missed and puts them back in the right order.

How it works

Realtime workspace. The console keeps a live connection to the chat platform and shows every active conversation at once, each with the visitor's context alongside the thread.

Wait reminders. Each conversation tracks how long the visitor has been waiting for an agent reply. Reminders step up as that time grows, so the most neglected conversation is always the most visible one.

Reply suggestions. For the conversation in front of them, the agent can ask for an AI-drafted reply. Nothing is sent automatically: the agent approves, edits or discards the suggestion, and rates it, so the quality of suggestions can be judged over time from real use.

Wake recovery. When the machine wakes from sleep, the app detects the gap, re-establishes its connection, asks the platform for the messages that arrived while it was asleep, and merges them into each conversation in their original order, so the agent sees a complete thread instead of a silent hole.

Technical challenges

A laptop that sleeps drops its realtime connection without any clean signal, so the app has to recognise the gap itself rather than wait to be told.

Filling in missed messages across many open conversations, in the right order, without showing anything twice.

Keeping an agent's attention on the conversations that need it most while many are open.

Keeping the AI assistance useful without letting it speak for the agent: suggestions are drafts, and the agent's approval and rating stay in the loop.

Capabilities

Many concurrent chats in one realtime workspace
Visitor context beside each conversation
Escalating wait reminders
AI reply suggestions that the agent approves, edits and rates
Detection of sleep and wake, with missed messages filled back in automatically, in order

Technology

Impact

Agents could keep many conversations moving without losing track of who was waiting longest, reply faster with suggestions they stayed in charge of, and trust that a conversation on screen was complete after their laptop had been asleep. No metrics are claimed.

Skills demonstrated

Support workflow design
Working in a small development team