AI-Powered Customer Service Cockpit

Support teams handle a large number of recurring requests every day. Each message has to be read, classified, and answered, a repetitive, time-consuming process that quickly becomes a bottleneck at high volume and crowds out the focus on the genuinely tricky cases.

The idea: an assistant cockpit that automatically prepares incoming requests, delivers a short assessment, and proposes a draft reply, as a starting point for the team members, not a replacement. It should let them work faster and more consistently without taking the decision out of their hands.

I set two constraints as guardrails for the concept from the very beginning:

• Privacy first: inference runs on self-controlled infrastructure instead of handing content to a third-party LLM provider
• Reliability over false confidence: the AI assists but never invents facts, and every reply stays in human hands before it is sent

To show that the concept holds up, I built it as a prototype: a cohesive system in which team members work from request to send-off in a single, reactive interface. Three guiding ideas shape the architecture.

The prototype shows that the pattern holds up end to end. Routine work is taken off the team's plate, keeping them free for the demanding cases:

• Automatic preparation of every request with a compact assessment and a draft reply
• Editable drafts, multilingual and in different tones
• Self-controlled inference (locally via MLX, scalable via vLLM on GPU), no third-party LLM provider, privacy-friendly
• Reactive interface with live updates instead of waiting for finished results
• Live translation in both directions: each side reads and writes in its own language, the original remains the source of truth
• Human-in-the-loop: traceable approval and adjustment of every reply

The crucial part remains the interplay of human and machine: the AI delivers a good starting point, while responsibility for the final reply stays with the human.