Ingestion & normalisation
Helpdesk webhooks land on a FastAPI endpoint, are normalised into one ticket shape regardless of channel, and queue in Redis. The queue is the backpressure valve: peak season floods the queue, not the model provider.
Classify every ticket, answer the routine ones from cited policy, route the rest with a draft attached.
Reference engagements illustrate how Elicon would approach a project of this kind — the problem, the architecture and the decisions. They are examples, not accounts of past client work.
THE SITUATION
A multi-brand retailer whose support volume grows faster than it can hire. The engagement adds a triage and answering layer that classifies every inbound ticket, answers routine questions with cited product and policy passages, and routes everything else to the right person with a drafted reply attached.
Support volume outgrows the team, first responses slip, and routing depends on one senior person reading every ticket.
APPROACH
ARCHITECTURE
Helpdesk webhooks land on a FastAPI endpoint, are normalised into one ticket shape regardless of channel, and queue in Redis. The queue is the backpressure valve: peak season floods the queue, not the model provider.
A fine-tuned small model assigns intent, brand, urgency and language with a confidence score per label. Anything under threshold falls through to a general queue — a wrong confident route costs more than an unrouted ticket.
Catalogue, policies and resolved tickets indexed in pgvector and filtered by brand and country at query time, so one market's policy can never be cited to another's customer. Hybrid keyword and vector retrieval, because SKU codes are keyword problems.
Above the auto-send threshold, a cited answer goes out. Below it, the same draft waits on the ticket for a person to approve or rewrite. Nothing sits in an ambiguous middle state.
DECISIONS
Routing is narrow, high-volume and low-creativity, with labelled history already in the helpdesk. A small model is cheaper per call and much faster.
TRADE-OFFIt needs retraining as the catalogue changes, so the retraining loop is built as part of the system rather than left as a script.
A good automatic answer saves minutes; a bad one costs a customer relationship.
TRADE-OFFA lower automation rate at the start, raised gradually against real rejection data rather than set optimistically on day one.
The team already works in it and reporting is built on it. Replacing it would turn an AI project into a migration.
TRADE-OFFSync latency and the helpdesk's API limits become part of the design, so the queue and idempotent writes carry more weight.
STACK
Work of this shape usually runs as AI Product Launch →
SIMILAR PROBLEM?
One conversation with an engineer is usually enough to tell whether this is a few weeks of work, a few months, or something you shouldn't build at all.
Talk to an engineer →