Logistics Company
2026 – Present
Discovery phaseDiscovery phase run as a proof of concept on real data: AI-assisted handling of a logistics company's customer service emails.
Overview
A logistics company (3PL) wanted to know how far AI could go in its customer service mailbox, where representatives handle orders, receipts and stock questions by hand. I ran the discovery phase as a proof of concept on real data, from ingesting the mailbox to a production architecture. Along the way, the solution took a shape the original plan didn't include: a Chrome side panel right next to Gmail.
Key Contributions
Data ingestion and cleaning
Read-only ingestion of the mailbox through the Gmail API — incremental, resumable, raw data kept untouched.
Each email split into author text, quoted replies, signatures and forwards without discarding anything, and real documents separated from logos and tracking pixels.
Evaluation harness and benchmarks
An evaluation harness on real emails: curated eval set, versioned prompts, expected outputs and an LLM judge.
Anthropic, OpenAI and Google models compared on quality and cost; the target runs on the Claude API.
Four ways of reading documents compared: PDF text layer, OCR, native model input and Extend AI.
Scope and end-to-end solution
Requirements drawn from the mailbox itself, and a first release scoped to verified senders and four request types.
A Chrome side panel next to Gmail shows representatives what the system understood; they stay in control, and nothing happens automatically.
Integration and architecture
WMS integration designed around the existing EDI channel: the model never writes to the WMS, and code acts only after a representative does.
A production architecture: eleven agents orchestrated by code, per-customer memory, evaluation in CI and a kill switch, on Google Cloud.
Internal mail in the mailbox, set aside before any AI call
Attachments that weren't business documents
Cost to triage an email with Claude Haiku 4.5