TRACE — Resolution Intelligence Agent

Inspiration

Every support team has gotten this message: "It's been two days, this is my third follow-up — can someone just tell me what's happening with my ticket?" You open the ticket to answer, and it says "In progress." That's it. So you scroll through forty replies, work out who has it now, who had it before, when it moved to another team and why it came back — just to write one honest sentence back.

Helpdesk have dozens of features — dashboards, SLA timers, reports, AI summaries — but none of them answer the three things you actually need to resolve a ticket fast: what's wrong, where it's stuck, and why. They tell you a ticket is late, not why it's late

What it does

TRACE is a resolution-intelligence agent. It reconstructs what actually happened to a ticket — we call it a resolution graph — from the replies, reassignments, and status changes already in the ticket, and then reasons over that structure. Nothing is guessed; every number is computed from the graph.

• Time ledger — every hour of the ticket attributed to a team and marked active-work vs. waiting. (Our demo ticket: 52h old, 35h of it waiting — 67% — with nobody working it.)
• Explain the delay — the costliest handoff, information asked for twice, and who is blocking the ticket right now with SLA remaining.
• Handoff reasons — each team-to-team move records why it moved, how long the team held it, and what was missing on arrival.
• Ask TRACE — natural-language Q&A answered strictly from the graph; it refuses to invent numbers.
• Handoff guardian — before a ticket moves to the next team, TRACE checks it against what that team needed on past tickets and holds the handoff if a required input is missing — even drafting the request.
• Proactive alert — watches open dependencies and escalates to the accountable owner once a wait runs longer than normal and the SLA is at risk.
• Manager view — the same graph across many tickets, turning "tickets are slow" into a specific, fixable process problem.

Accomplishments that we're proud of

• A working engine that turns a raw ticket into a resolution graph and reproduces the full story end-to-end — 52h total, 35h waiting, four handoffs, the exact blocker.
• It computes structure first, then reasons — so it answers "who is waiting on whom, and for how long," which no summarizer can.
• It runs both backward (explain what went wrong) and forward (predict and prevent the next bad handoff) from one engine.
• Grounded answers that never fabricate a number.

What we learned

• The valuable signal in a ticket isn't the text — it's the structure of custody over time. Once you compute that, explanation, prediction, and prevention all fall out of the same model.
• "Late" is easy; "why it's late" is the hard, useful part — and it's exactly what existing tooling skips.
• Grounding an LLM in a computed structure is far more reliable than asking it to summarize a thread.
• Most of a multi-team ticket's life is invisible waiting — and it's largely recoverable by fixing one or two handoffs.

Built With

• css
• fdk
• freshdesk
• html
• javascript
• llm
• mcp
• node.js