Smasher — AI Support Assistant
The conversations shown are demo tickets on test accounts (Om Nom, Bonkers) — not real player conversations.
About This Project
A retrieval-augmented support assistant wired into Helpshift: it reads incoming tickets, searches the public help-center knowledge base, and either answers the player directly or drafts a private note for agents — depending on how confident it is.
Most support volume is repetitive questions already answered in the help center. Agents spent time re-typing known answers instead of handling cases that need judgment.
Together with Helpshift automations, repetitive ticket volume dropped ~30% while CSAT held at 4.5+/5. Shadow mode made the rollout safe: the assistant earned trust as an agent tool before answering players directly.
RAG assistant on top of Helpshift that answers players from the public knowledge base — part of cutting repetitive support volume ~30% at 4.5+/5 CSAT.
What I Worked On
System
Helpshift webhook → FastAPI service → RAG over the public help-center knowledge base: articles are scraped, chunked, deduplicated, and embedded into a searchable index.
Confidence-scored answers with two modes: high-confidence replies go to the player (live mode); low-confidence drafts are posted as private notes for agent review (shadow mode).
Escalation built in: below-threshold or sensitive tickets get tagged for humans, and Helpshift automation routes them to the right queue with the reason attached.
Conversation history for context-aware replies; automatic language detection with English and Russian knowledge bases.
Answers in under ~3 seconds, with sourced help-center links in every reply.