Project Story: The AI-Powered Freight Assistant

Inspiration

I currently work as a Product Manager at a freight forwarder, where daily operations rely heavily on Operations Specialists (OPs). An OP handles the entire lifecycle of a shipment—from container loading to customs clearance and trucking. This process relies heavily on manual email communication, which is time-consuming and prone to human error.

In this industry, errors translate directly into financial losses. For example:

• Arrival Notice Delays: Can lead to port demurrage fees ranging from $300 to $500 per container per day.
• Late Customs Filings: Can result in fines as high as $5,000.

The primary revenue model for freight forwarders is based on service fees and commissions. This is directly linked to OP performance. Even the most senior OPs average between $7,000 and $10,000 in annual fines. This product was built to improve this high-risk, human-dependent profit model.

What it does

Imagine a 24/7 digital OP. Upon receiving an email, the agent analyzes the intent based on pre-defined "agent skills." It parses the subject, body, and attachments to take action:

1. New Orders: Automatically creates a shipment in the internal system based on attached files.
2. Carrier Arrival Notices: Instantly generates and sends the forwarder’s own Arrival Notice to the consignee (preventing those late fees!).
3. Customs Coordination: Automatically syncs data with customs brokers based on client documentation.

How I built it

I started by mapping out the workflows for different shipment types (Import FCL, Import LCL, etc.). I then designed a Permission and Responsibility Hierarchy, separating actions into "Auto-executable" and "Requires Approval" to ensure risk control.

The architecture positions the AI as a "Senior OP Assistant." I chose Gemini 3 for three specific reasons:

• Multimodal Context: It processes email text, PDFs, and images simultaneously.
• Long Context Reasoning: I can feed the entire domain playbook (e.g., MBL vs. HBL rules) into the system prompt.
• Structured JSON Output: It supports strict schema constraints, ensuring responses map to executable business logic.

Every output must follow a defined schema: $$Action \in {CREATE, UPDATE, ESCALATE, APPROVE}$$

Challenges we ran into

1. Domain Ambiguity: Documents often contain both Master Bill of Lading (MBL) and House Bill of Lading (HBL) data. Distinguishing between them requires deep logic, not just field extraction.
2. Alignment Issues: A single email might contain one PDF that covers multiple containers or several HBLs. The system must accurately "de-batch" these.
3. Automation vs. Risk: We intentionally kept $20\%$ of the process—the high-risk "last mile"—under human control to protect the business.
4. The "Demo vs. Reality" Gap: Real-world info isn't always in the email (e.g., a phone call update). I added a manual "Edit" function to allow OPs to patch information manually.

Accomplishments that we're proud of

• Accuracy: The model maintains a 95% confidence level in shipment identification.
• Efficiency: We reduced the average workflow time by 40%, with complex scenarios seeing improvements of up to 55%.
• User Feedback: Our first-line OPs gave us an NPS of 9. Management gave an 8, focusing on commercial scaling and liability.

What I learned

I learned to integrate business and profitability logic early. I calculated the unit economics from day one:

• LLM API Cost: $\approx \$0.01$ per shipment.
• Estimated Total Cost: $\approx \$2.00 - \$3.00$ per shipment (including 3rd-party APIs).

This led me to a "Charge per Shipment" model rather than a subscription, aligning costs directly with the value provided.

What's next for Freightbot

• Product Roadmap: Support more shipment types and implement an "Agent Browser" function for various FMS.
• Business: Move to GTM (Go-To-Market). I will start by onboarding our internal OP teams before expanding to external vendors.

Built With

• css
• gemini-api
• javascript
• pdf-parse
• react
• vite