Independent Portfolio Project · Solutions Engineering

Freight Operations AI
POD Recovery Assistant

A lean, interview-ready build for a simulated presales engagement in freight brokerage operations: an AI-assisted workflow that recovers missing proof-of-delivery documents without becoming an autonomous agent.

The problem

Summit Freight Partners is a fictional, mid-market non-asset freight brokerage operating across North America. Operators manually chase proof-of-delivery documents across email, text, and phone. Missing or mismatched documents delay billing and create repetitive, low-value work for the team.

Every number below is an illustrative assumption to validate in real discovery, not customer evidence.

The solution concept

One narrow workflow, deliberately scoped: a controlled process with an AI-assisted decision step, not a general autonomous agent platform. That restraint is part of the technical story.

1. 01TMS marks a load delivered.
2. 02The workflow checks whether a usable POD exists after the configured wait period.
3. 03Policy and load data determine whether outreach is eligible and which contact and channel are allowed.
4. 04The AI drafts a grounded action with SOP citations; the demo requires approval before sending.
5. 05A simulated carrier response returns a document.
6. 06The system extracts candidate identifiers, then deterministic rules validate the match.
7. 07A valid match updates the simulated TMS; ambiguous cases create a human-review exception.
8. 08Every decision, approval, external action, and result enters the audit timeline.

Lean architecture and boundaries

Say every limitation plainly. The project demonstrates reasoning and workflow design, not production certification.

AI decision contract and safeguards

The model may recommend — the application decides. It is only ever given the selected load record, approved contact data, relevant SOP sections, and the allowed action list.

• Use only supplied data; never invent a load, contact, status, document fact, or policy.
• If a required field or governing rule is missing, return needs_review.
• Return structured output that the application validates against an allowlist.
• Do not let the model directly send messages, update the TMS, or approve documents.
• Require human approval for the portfolio demo’s external-action step.
• Use deterministic rules to validate load number, carrier, delivery date, and duplicate document state.
• Record model version, prompt template version, source IDs, output, validation result, and operator decision.

A compelling AI demo shows when the system refuses or escalates, not only when it succeeds.

How this will be demonstrated

The finished deliverable pairs a seven-minute customer-facing demo (problem, workflow, a run exception, and value framing) with a twelve-minute technical walkthrough (architecture, state and audit, the AI contract, integration boundaries, and the production path). It ships alongside the presales artifacts that support a real proof-of-value: a discovery brief, a gap log, a security brief, and a two-page POV plan.