AI CUSTOMER WORKFLOW
A serverless agent that turns a new support ticket into a Slack-ready handoff: related Linear issues, likely owning engineer, draft customer reply, and next steps.
← Back to ProjectsThis anonymized portfolio version is based on a production support workflow: CSMs receive customer tickets and need to know whether a related engineering issue already exists, who owns it, and what they can safely tell the customer.
The agent receives a Pylon ticket-created webhook, uses Claude with tool use to search Linear, forces a structured triage result, and posts the final handoff to Slack.
Portfolio focus: Customer workflow discovery, LLM tool use, structured outputs, operational guardrails, and integration design across support, engineering, and Slack.
A customer issue arrives with title, body, priority, tags, and account context.
The webhook is validated, normalized, and sent into the triage loop.
The model searches Linear with ticket-specific queries and gathers related work.
The loop ends only when a typed triage result is submitted through a tool.
Support gets the summary, likely owner, draft reply, and follow-up actions.
New ticket — Northwind Logistics
Users hitting a redirect loop after Okta login
Symptoms align with a known in-progress SAML callback bug where RelayState can redirect users back to the login page.
Priya Raman
The problem is not “add AI.” It is the repeated first-pass investigation a CSM does before they can confidently route a customer issue.
The agent compresses that lookup into a reviewable handoff instead of hiding the decision.
The repo includes the Lambda handler, Linear GraphQL client, Slack formatter, local fixture runner, FastAPI Linear mock, tests, and an AWS SAM template.
The demo path keeps the code reviewable without requiring real Pylon, Linear, or Slack accounts.
Tool use for Linear search. Claude can run multiple searches with different terms when the first query misses the real issue.
Forced structured output. The final result comes through a submit_triage tool, avoiding fragile parsing of free-form model text.
No invented tickets. Related issue IDs are joined back to the tool-call history before they appear in the final result.
Bounded cost and latency. The loop caps turns, caches the system prompt, and falls back to stdout when Slack is not configured locally.