70%
faster lookup
Median time to find an answer, measured in agent screen recordings.
Sample project · Paperkite
Sample project: an illustration of how we approach this kind of work, not a real client engagement.
Industry
AI · Retrieval-based support assistant
Timeline
10 weeks
Team
1–2 senior engineers
Stack
01 · The challenge
Paperkite’s support agents searched six tools to answer one customer question: a wiki, two folders of policy PDFs, a ticket archive, release notes and a chat history. New hires needed months before they could find answers quickly.
An earlier chatbot experiment had been shelved because it answered confidently from outdated policies and could not show where an answer came from.
02 · Our solution
We indexed policies, resolved tickets and release notes with team-level access rules, and built a retrieval pipeline that returns an answer only when supporting sources are found. Every answer carries citations that open the original document at the right paragraph.
Documents carry owners and review dates, so stale content is flagged and ranked lower. An evaluation set of 400 real questions ran on every change, which let the team improve answer quality without guessing.
Sample figures for this illustrative case study, measured against the period before launch.
70%
Median time to find an answer, measured in agent screen recordings.
2,000
Weekly active agents within three months of rollout.
93%
Answers backed by at least one source the agent could open.
−35%
Time for a new agent to reach the average handling-time target.
Education Sample project · Tandem Learning
Flutter app with offline lessons, streaks and teacher dashboards across iOS and Android.
Healthcare Sample project · Northwind Health
Northwind ran consultations over phone calls and shared spreadsheets. We built a booking, video and e-prescription platform that now handles 9,000 visits a month across three countries.
Retail Sample project · Lumen Retail
Headless commerce rebuild with localised pricing, search and checkout that cut cart abandonment.