DocFlow AI
Self-improving AI agent that classifies and extracts data from customs documents in real-time
About This Project
DocFlow AI is an autonomous AI agent that automates international trade document processing end-to-end. It ingests emails via Gmail webhooks, classifies attachments (import declarations, bills of lading, invoices, tax documents) using multi-model AI on GMI Cloud's NVIDIA-powered infrastructure, and extracts structured data with type-specific schemas.
What makes it unique: a self-improving prompt evolution system. When users report classification errors, the agent collects the mistakes, analyzes patterns with a meta-prompt, and generates additive rules — never removing existing knowledge. New prompt versions are created but require human approval before activation, ensuring zero regression.
The pipeline includes smart retry (re-queries AI for missing fields with focused prompts), 3-tier email grouping (thread ID → normalized subject → new process), transient error recovery with progressive backoff, and stuck process auto-healing.
Built for Brazilian customs (DI, BL, NF, ICMS), the architecture is document-agnostic — the same agent can be trained on US customs forms (CBP 3461/7501, Commercial Invoices, HTS classifications) by simply providing examples.
Key features:
- Real-time email ingestion via Gmail Pub/Sub
- Multi-model AI classification & extraction via GMI Cloud (NVIDIA GPUs)
- Self-improving prompts with versioning, rollback, and accuracy tracking
- Human-in-the-loop review dashboard
- Processing in under 30 seconds per document
Tracks
Built With
Repository
Submitted March 28, 2026 at 3:45 PM