Last-Mile Delivery Exception Agents
A LangGraph multi-agent system that reads messy delivery logs, picks a playbook-grounded action, escalates by policy and writes the customer message.
| Shipment | Action | Escalated | Tone | Task | Escalation | Tools | Coherence | Latency | Pages |
|---|---|---|---|---|---|---|---|---|---|
| SHP-001 | N/A | NO | N/A | PASS | N/A | PASS | 5/5 | 0.0s | noise |
| SHP-002 | RESCHEDULE | YES | FORMAL | PASS | PASS | PASS | 4/5 | 8.4s | p5, p5, p9, p2 |
| SHP-003 | RESCHEDULE | NO | CASUAL | PASS | PASS | PASS | 5/5 | 8.4s | p3, p3, p3, p2 |
| SHP-004 | REPLACE | YES | FORMAL | PASS | PASS | PASS | 5/5 | 8.9s | p4, p4, p4, p6 |
| SHP-005 | REROUTE_TO_LOCKER | YES | CASUAL | PASS | PASS | PASS | 5/5 | 10.0s | p2, p2, p3, p10 |
| SHP-006 | RETURN_TO_SENDER | NO | CASUAL | PASS | PASS | PASS | 5/5 | 12.4s | p5, p5, p5, p2 |
| SHP-007 | REPLACE | YES | FORMAL | PASS | PASS | PASS | 5/5 | 10.3s | p6, p6, p6, p6 |
| SHP-008 | RESCHEDULE | YES | CASUAL | PASS | PASS | PASS | 5/5 | 10.9s | p2, p3, p3, p5 |
| SHP-009 | REPLACE | YES | FORMAL | PASS | PASS | PASS | 5/5 | 9.3s | p6, p4, p4, p3 |
| SHP-010 | N/A | NO | N/A | PASS | N/A | PASS | 5/5 | 0.0s | noise |
Proof-of-concept results on 10 curated shipments; not a production accuracy estimate.
SQLite (12 customers, 6 lockers), 2 CSVs, 10-page playbook in 36 chunks; gpt-4o-mini agents, gpt-4o critic.
Log reader, customer lookup (PII-gated), locker eligibility, playbook RAG search, deterministic escalation rules.
Preprocessor, orchestrator, Resolution, Critic (x2), Communication, Finalize; prompts, routing and 4-layer escalation.
All 10 shipments with traces, page citations and the five metrics, plus 3-run tuning history.
Shadow-mode pilot, rule engine as escalation authority, larger held-out test set.
34-slide deck in the course template; every rubric section covered, appendix with prompts, limits and cropped notebook screenshots.