Resolvr — Autonomous Support Resolution Agent
Problem
Support teams drown in repetitive tier-1 tickets — password resets, invoices, how-tos. Hiring to keep up is expensive, queues blow out, and naive auto-replies hallucinate policy or promise refunds no one approved.
What the agent does
A standalone full-stack product (FastAPI + React) that takes a raw support ticket to a finished outcome on its own: it classifies the request, retrieves the right knowledge-base articles by semantic search, decides whether it can safely resolve or must escalate, and drafts the reply grounded strictly in the KB — never inventing policy and never promising a refund.
Autonomy
A four-tool loop — classify_ticket → search_kb → decide_action → draft_resolution. Retrieval is real RAG (Ollama embeddings + cosine); the reply is written by a local Ollama model (qwen2.5:14b), falling back to a cloud API key, then to a template, so it self-hosts at zero per-token cost. A hard safety gate forces escalation on security, legal, abuse, and refund tickets — proven at 100% must-escalate recall by the pytest eval suite. Try the lite in-browser demo below; the full product runs separately.
- Must-escalate recall
- 100%
- LLM backend
- Ollama + API
- Stack
- FastAPI + React