Case study

High-Load Support Assistant

A grounded support assistant answering catalogue, delivery and returns questions at sale-day traffic — with sub-ten-second worst-case responses and no invented policy.

Client type: Online fashion retailer with seasonal traffic peaks (sample scenario)
Industry: E-commerce
Duration: 5 months

Sample contentSample engagement. Illustrative scenario used while the site is in build — not a named client project.

e commerce2026

Python
FastAPI
PostgreSQL
LLM integrations

01Challenge

Support volume tripled during sale periods, and most of it was repetitive: where is my order, does this size run small, what is the returns window. Generic chatbot products either answered from nothing or invented policies the retailer then had to honour.

02Solution

A retrieval-grounded assistant built over the retailer's own catalogue, policy pages and order status API, with autoscaling sized against historical peak traffic and a strict rule that unanswerable questions escalate to a human.

03Business impact

The support team stopped spending sale weeks on repetitive questions and could focus on the complaints and exceptions that actually needed judgement.

04Approach

01Content pipeline normalising catalogue and policy pages into a consistent structure
02Embeddings stored alongside the operational data rather than in a separate service
03Answer grounding enforced — no source, no answer
04Load testing against three seasons of historical peak traffic before launch
05Escalation path to a human agent with full conversation context attached

One stateful system, not two

Storing embeddings in the existing PostgreSQL cluster rather than adding a dedicated vector database removed an entire class of operational work — backups, failover and consistency were already solved.

Start a conversation

High-Load Support Assistant

One stateful system, not two

Services involved

More case studies

Similar problem on your side?Let's talk.