Build an AI System to Manage Logistics Vendor Performance

Many logistics SMBs rely on the vendor management modules within their TMS, like MercuryGate or Trimble TMW. These platforms provide basic scorecards for on-time performance and tender acceptance rates, but only if the data is entered manually and perfectly. The systems cannot automatically parse a carrier's PDF performance summary or extract transit time details from a confirmation email. This forces logistics coordinators to spend hours on data entry just to keep the scorecards current.

Consider a 15-person freight brokerage managing 40 carriers. One of their top carriers begins to have consistent unloading delays at a specific receiver, but the evidence is buried in unstructured emails from the dispatch team ('driver held up for 3 hours at XYZ warehouse'). Because these exceptions are not manually logged as 'late' in the TMS, the carrier's on-time delivery score remains at 98%. The brokerage only discovers the systemic issue after their client threatens to pull business due to repeated delays.

The structural problem is that TMS platforms are designed as transactional databases, not analytical engines. Their data models are rigid and expect clean, structured inputs for pre-defined fields like 'Pickup Time' and 'Delivery Time'. They have no native capability to apply natural language processing to an email thread or an attached PDF to identify root causes for delays. The architecture is fundamentally unequipped to handle the messy, real-world data where performance signals actually live.

As a result, decisions are made on incomplete data. The brokerage continues to award loads to the underperforming carrier while potentially penalizing a reliable carrier whose only fault was a single, well-documented mechanical failure. The lack of a unified, intelligent view of performance introduces significant operational risk and damages client relationships.

The engagement would begin with a complete audit of your carrier data sources. Syntora would map out every place performance data lives: your TMS database, carrier web portals, shared email inboxes, and folders of PDF reports. You would receive a data map showing exactly what can be automated and what signals can be extracted.

The core of the system would be a Python service using the Claude API to parse unstructured text from emails and documents. Claude API is chosen for its ability to extract JSON-formatted data from complex text, turning a carrier's email update into structured fields like `delay_reason` and `delay_hours`. This FastAPI service processes each document in under 2 seconds, handling over 500 carrier updates per day. All extracted data, along with 12 months of historical TMS data, is stored in a Supabase Postgres database.

The final deliverable is not another software login for your team. The processed performance data feeds a web-based dashboard on Vercel that highlights trends and outliers. Key alerts can be sent directly to Slack or email. You receive the complete Python source code and runbook, and the system runs in your own AWS account on Lambda for under $50 per month, ensuring total data ownership.