AI · LLM integration
AI support assistant
Meridian Software — 2025
A retrieval-augmented assistant trained on product docs, wired into an existing helpdesk with human handoff.
- Result
- 62% of tickets resolved without an agent
- Timeline
- 10 weeks
The challenge
A B2B software company's support queue grew faster than hiring. Most tickets were repeat questions already answered somewhere in a 900-page documentation set.
What we did
- Indexed documentation, changelogs and past resolved tickets into a vector store.
- Built a retrieval pipeline with citations so every answer links to a source page.
- Added confidence thresholds that escalate to a human agent instead of guessing.
- Instrumented every conversation for accuracy review and prompt tuning.
The outcome
- 62% of incoming tickets resolved without agent involvement.
- First-response time dropped from hours to seconds.
- Support headcount reallocated to onboarding instead of triage.
Stack
- Python
- FastAPI
- PostgreSQL + pgvector
- OpenAI
- Zendesk API
It answers like someone who actually read our docs, and it knows when to hand over. That trust was the whole battle.
Tell us about your project
Free consultation, no obligations. Send the details or schedule a call for the fastest reply.