I built the Instamart Copilot Part C, an automated support‑ticket triage pipeline that ingests tickets, redacts PII, classifies them with a mock LLM, handles guardrails, retries transient errors, and routes tickets based on priority and the premium‑customer segment from Part B.
I loaded 164 tickets from data/raw/tickets.csv.
Exactly 4 tickets (T0157–T0160) triggered the empty‑ticket guardrail and were recorded with status skipped_empty, category "", and tokens_used = 0.
Exactly 4 tickets (T0161–T0164) triggered the payload‑length guardrail and were recorded with status flagged_long, category "", and tokens_used = 0.
PII redaction performed 278 substitutions (55 emails, 91 phones, 132 order IDs).
Exactly 19 transient failures were encountered during the full evaluation run, and all 19 were successfully recovered on the second attempt without exceeding 3 attempts.
Nine tickets produced disallowed categories and were flagged with status flagged_invalid_output.
Under the few‑shot prompt I achieved 147 correct classifications out of 147 classified tickets.
Under the zero‑shot prompt I achieved 126 correct out of 147 classified tickets → 85.71 % classified accuracy.
Routing results: support_queue (NORMAL) = 102 tickets, security_team (HIGH) = 34 tickets, premium_desk (HIGH) = 11 tickets, human_review (NORMAL) = 17 tickets.
Official run token consumption was 21,568 tokens out of a 100,000‑token budget, leaving 78,432 tokens remaining.