Second Read

Inspiration

Every support dashboard tracks how fast tickets close. None of them ask if they should have stayed open.

We noticed that "resolved" is treated as a finished state, but in reality, a meaningful share of resolved tickets are silent failures — customers who stopped replying without confirming a fix worked, issues that quietly recur under different wording, or fixes that were never actually verified. No one goes back to check.

What it does

Second Read is a multi-agent AI system that re-audits closed support tickets to catch false resolutions before they resurface as bigger problems.

• Verifier Agent re-reads the closing conversation and judges whether the ticket was genuinely resolved, or silently dropped
• Pattern Agent cross-references the ticket against historical cases to catch disguised repeats — the same root cause hiding under different symptoms
• Reporting Agent synthesizes both findings into a manager-facing brief with a confidence score, occurrence count, and estimated cost impact

The result: a weekly briefing that tells a support team exactly which "resolved" tickets are worth a second look — and why.

How we built it

The frontend is a React + Tailwind interface (Dashboard, Cluster Map, Ticket Detail, and Manager Briefing screens), prototyped in Figma and exported to code. The backend runs a LangGraph pipeline chaining the three agents together, powered by Llama 3.3 70B via Groq. Ticket ingestion is designed around a Freshservice/Freshdesk MCP connector, with agents orchestrated in a pattern compatible with Freshworks Agent Studio.

Challenges we ran into

Getting multi-agent orchestration to feel genuinely useful rather than just a demo pipeline — we focused on making sure each agent's output (confidence score, pattern match, cost estimate) was something a manager could act on immediately, not just an AI-generated summary.

What's next

If shortlisted, we'd connect Second Read to live Freshservice data via MCP, build out the feedback loop so manager verdicts actually improve future flagging accuracy, and add real-time cluster detection instead of the current periodic audit.

Built With

• fastapi
• figma
• groq
• langgraph
• llama-3.3
• mcp
• python
• react
• tailwind-css
• typescript