The Client
A leading logistics company offering fast, reliable courier and transport services across B2B, B2C, and cross-border business segments, this organization manages a high-volume shipment network with stringent delivery SLAs and complex billing processes tied to Proof of Delivery verification. The business required data and AI capabilities that could extend operational intelligence beyond specialist data teams to frontline operations and finance.
The Challenge
Disconnected datasets and restrictive access to analytics prevent operations teams from making data-driven decisions that could optimize lane-level profitability. When manual document processing and lack of self-service capability bottleneck the workflow, the business suffers from delayed billing and reduced operational throughput.
- Analytics Bottleneck: Amazon Redshift was the primary analytics warehouse, accessible only to specialist data teams via SQL, leaving operations and finance teams without self-service access to AWB profitability, lane margins, and cost driver data.
- Manual POD Verification: Proof of Delivery processing was entirely manual, agents reviewed physical and scanned documents to extract fields, validate against shipment records, and flag exceptions, creating delays in billing closure.
- Siloed Operational Data: Three separate datasets, AWB Journey, Profitability/Revenue, and Trip-level files, existed in CSV and Parquet format without a unified analytics layer, preventing end-to-end operational analysis.
- Governance and Audit Gaps: No centralized access control, audit logging, or query traceability existed across analytics or document processing workflows, creating compliance and operational risk.
The Impact
- 300 PODs processed in pilot
- 80% OCR extraction accuracy target achieved
- Self-Service Business users querying operations data in plain English
The Solution
Hoonartek built two AI-powered workstreams on GCP. For analytics, AWB Journey, Profitability/Revenue, and Trip-level files were ingested via GCS into BigQuery raw tables with automated data quality checks. Four curated BigQuery views, Margin, Lane Rank, Cost Breakdown, and RCA Support, were built as the sole agent query surface. A BigQuery Conversational Agent translates plain-English questions into SQL, with results and RCA narratives governed by defined system guardrails. All agent queries are logged to Cloud Logging and BigQuery Job Logs, with IAM-controlled access restricted to curated views only.
For POD verification, Hoonartek deployed a Document AI and OCR-powered agent that enforces an image quality gate at ingestion (minimum 300 DPI), extracts key fields including AWB number, delivery date, receiver, weight, dimensions, and handwritten remarks, and validates the AWB against the ShippingID master via exact string match. Per-field confidence scores are assigned, with records below threshold routed to a structured manual review queue. Processing dashboards, audit logs, and downstream billing integration provide end-to-end traceability across the POD lifecycle. Hoonartek’s RealizeAI Framework, AgentOps, and ClarityX accelerators were deployed across both workstreams.
Key Benefits
- Self-Service Analytics: Business users can now query AWB profitability, lane margins, and cost drivers in plain English, no SQL skills or data team dependency required.
- Automated POD Processing: OCR-driven extraction and confidence scoring handles 300 PODs processed in the pilot with an 80% OCR accuracy target, accelerating billing closure and reducing processing error rates.
- Lane-Level Operational Intelligence: RCA narratives and profitability trend views surface root causes of lane-level losses directly to operations teams in real time, delivering 35% faster margin insights.
- Governed Query Execution: Full IAM access control, Cloud Logging, and BigQuery Job Log auditability ensure every agent query is traceable and compliant.
- Exception-Driven Operations: Only genuinely ambiguous POD records route to human reviewers, structured exception queues reduce operational backlogs and improve throughput.