// AI CAPSTONE ◆
The goal wasn't a perfect AI system. It was learning where to look for flaws.
For the capstone project of my Post Graduate Program in AI Agents for Business Applications (University of Texas, 2026), I built a proof-of-concept multi-agent workflow in n8n. Working from a high-level project guide, I configured each step of the workflow, adjusted parameters, wrote the agents' prompts, incorporated guardrails and evaluated performance across 10 shipment scenarios.
THE SCENARIO · LAST-MILE DELIVERY EXCEPTIONS
When a delivery goes wrong, like a missed attempt or an unavailable locker, someone has to review the shipment, choose a fix and tell the customer. The proof of concept explored how AI agents could help while following business rules and keeping a human in the loop.
◆ THE REAL OBJECTIVE
The goal wasn't to build a production-ready system. It was to evaluate what worked, identify weaknesses and determine what needed improvement before real-world deployment.
HOW THE WORKFLOW WORKS
STEP 01
Guardrails & rules
Screens for malicious inputs, filters routine events and flags cases that need a person.
STEP 02
Resolution agent
Recommends a fix from shipment data, locker availability and an exception playbook it searches with RAG.
STEP 03
Resolution critic
A second agent reviews the recommendation and accepts it or sends it back.
STEP 04
Communication agent
Drafts the customer-facing message explaining what happened and what's next.
STEP 05
Communication critic
Reviews the message for quality and tone before it goes out.
◆ HUMAN ESCALATIONAmbiguous or policy-sensitive cases are routed to a person instead of being decided by the agents.
EVALUATION · 10 SHIPMENT TEST CASES
◆ WHY THE SCORES LOOK LIKE THIS
I deliberately didn't fine-tune the workflow to chase better scores.
Tool-call accuracy reached 100% across the evaluated test cases, while the other scores revealed where the workflow's decisions needed further refinement.
WHAT I LEARNED
- 01
A workflow that runs isn't a workflow that's right.
The agents completed every tool call correctly, but their decisions didn't always follow business policy.
- 02
Guardrails and human oversight are essential.
Deterministic rules and clear escalation paths limit the risk of letting AI make unrestricted decisions.
- 03
Evaluation catches what a demo hides.
Testing exposed weaknesses I'd never have seen just by watching the workflow finish.
- 04
Responsible deployment takes refinement.
Before production, it would need broader testing, clearer decision policies, better prompts and ongoing monitoring.
BEFORE THIS GOES LIVE: MY RECOMMENDATIONS
- ☐Refine business rules, prompts and escalation criteria
- ☐Expand testing to more delivery scenarios and edge cases
- ☐Monitor tool failures, agent revisions and human overrides
- ☐Introduce automation gradually, starting with lower-risk cases
- ☐Keep humans on ambiguous or policy-sensitive calls
This is where my project management and implementation background meets AI: testing, governance, rollout planning and change management.