V2 · AI Automation Agency Hub

v2 of my AI Ticket Triage Assistant is done — fully verified, no loose ends.

Quick context: v1 handled single-issue tickets fine, but couldn't handle tickets describing more than one problem at once (e.g. "VPN down AND can't access SAP"). It would just pick one and route on that alone.

What I built for v2:

Extended the extraction step to detect multi-issue tickets, not just classify single-issue ones
Added a new top-priority routing rule — any multi-issue ticket goes straight to Service Desk for human triage, overriding the normal category/priority logic
Ran regression tests against my original single-issue tickets (Okta SSO, Office WiFi, sticky-keyboard) to confirm nothing broke
Manually checked every wire on the canvas — all 5 Switch outputs correctly connected, with Service Desk properly receiving both the multi-issue path and the fallback path

The part I actually want feedback on: two backlog items surfaced naturally through this process, and I'm deliberately not building them today

Error handling for malformed Gemini output (what happens if extraction returns something unexpected)
A design smell — 4 near-duplicate Log nodes that should really collapse into a single Merge-then-Sheets-node structure

Leaving v3 alone for now and writing these down instead of chasing them immediately.

Anyone here dealt with #1 in production — how do you handle malformed structured output from an LLM node without the whole workflow breaking downstream?

Difference from the LinkedIn version: no hook engineering, leads straight with the technical summary, and the closing question is a real technical ask (error handling for malformed LLM output) rather than a reflective prompt — matches how the community actually engages, based on Raj and Chetan's past comments.

5 comments