Challenge
Atlas's support team was drowning in repetitive tickets. Response times had climbed past 20 hours and hiring couldn't keep up with customer growth.
Solution
- Built a retrieval pipeline over 40k help-center articles, tickets and changelogs using pgvector hybrid search.
- Created a copilot inside their helpdesk that drafts replies with inline citations agents can verify in one click.
- Set up an evaluation suite of 600 real tickets to measure accuracy before every prompt or model change.
- Added guardrails for PII, refunds and escalation so sensitive tickets always reach a human.
Result
First-response time dropped from 20 hours to 2, agents resolve 45% more tickets per day, and CSAT went up 12 points.
Tech stack
- Claude
- NestJS
- PostgreSQL
- pgvector
- Next.js
Services: AI & Automation · Backend & API Development
