OverC
AvailableStart a project

(05)Case study · AI Product

Customer Support · 2026

Sage Desk

A retrieval-augmented copilot drafting grounded answers from 30,000 help-center articles and past tickets.

Client
Sage Desk
Year
2026
Services
AI Product, RAG Pipeline, Evaluation
Stack
Python · FastAPI · Anthropic · pgvector · React · AWS
All works
46%Faster ticket resolution
92%Drafts accepted by agents
1.2sp95 time to first token

(01)Challenge

Support agents spent most of each ticket searching for the right answer across docs, macros and old conversations — and answers varied by who picked up the ticket.

(02)Approach

  1. 01

    Built a hybrid retrieval pipeline (vector + keyword) with source citations on every sentence.

  2. 02

    Created an evaluation harness of 1,200 real tickets to measure accuracy before every release.

  3. 03

    Added guardrails for tone, PII redaction and confident refusal when sources are missing.

(03)Outcome

Agents review and send instead of search and write. Answers are consistent, cited, and measurably more accurate release over release.

(04)Architecture

How it's built.

A simplified view of the production system, following a request from the client through to the data layer.

(→)Next case study

Kiln

A headless storefront with sub-second pages and a real-time 3D product configurator.

(05)Contact

Let's buildyours next.

Tell us about your product, timeline and constraints. You’ll hear back from a senior engineer within one business day — with questions, not a sales deck.