AI Copilot for Support Operations

Context

Dewaweb’s support team handles real volume, and most incoming questions are variations of things they’ve already answered hundreds of times — but that institutional knowledge was scattered across years of resolved tickets. The team wanted to put that knowledge in front of agents at the moment they need it, without taking humans out of the loop. Direct LLM-to-customer replies weren’t the goal at this stage; a copilot that suggests answers and lets the agent ship them with a click was.

What I built

I helped migrate the support stack to a more extensible platform, then built the copilot pipeline on top of it:

• Designed and implemented a Retrieval-Augmented Generation pipeline using Gemini’s file search store as the retrieval backend, grounded in Dewaweb’s own historical support corpus.
• Orchestrated end-to-end flows in n8n — message ingest, retrieval, suggestion generation, follow-up actions — all inspectable as nodes.
• Ingested and cleaned the historical ticket data so suggestions come from real precedent, not hallucinated policy.

The system surfaces a suggested reply next to each incoming ticket; the agent reviews, edits if needed, and sends. Full AI direct-reply is on the roadmap for later, once model behavior and tone are dependable enough — for now the bar is assist, don’t replace.

Stack

n8n · Gemini API · file search store · Python · webhook plumbing

Outcome

In production. Early agent feedback puts the suggested replies at roughly 90% helpful for the cases they cover — measurable speedup on the high-volume repetitive tickets, with humans still on the trigger. Full LLM replies remain a future milestone; the copilot is the bridge.