Custom CRM integration with LLM, RAG, and CAG systems for a major Chinese agricultural exporter, featuring intelligent data querying, automated deal creation, and OCR-powered document verification.

Python · LangChain · QwenVL 3 · PostgreSQL · FastAPI · RAG · CAG · OCR · Claude · Redis
Forward East Group is one of China's largest agricultural importers, with over 20 years of experience supplying vegetable oils, feed additives, grains, and oilseeds from Russia, the EU, Latin America, and ASEAN countries to the Chinese market. With 4+ billion yuan in annual trade volume and partnerships with 300+ foreign enterprises, their operations generate massive amounts of data across deals, clients, suppliers, shipping documentation, and quality certifications.
The company needed to modernize their custom CRM to handle the complexity of international agricultural trade. We implemented an AI-powered intelligence layer that allows staff to query their entire business history in natural language, automates the creation of deals and client records, and uses vision AI to verify export documentation for inconsistencies before shipments clear customs.
The project was delivered in three stages over four months, building the data intelligence layer first, then implementing automation agents, and finally deploying the document verification system.
| Stage | Focus Area | Status | Key Deliverables |
|---|---|---|---|
| 1 | CRM Intelligence Layer | Completed | RAG and CAG integration with existing CRM, natural language querying interface, historical data indexing, multi-language support (Chinese, Russian, English) |
| 2 | Automation Agents | Completed | Deal creation agent, client onboarding agent, supplier management agent, task-specific tooling for each workflow |
| 3 | Document Verification | Completed | QwenVL 3 OCR pipeline, multi-document cross-referencing, inconsistency detection, export compliance checking |
Forward East's CRM contained two decades of trade data: deal histories, supplier contracts, shipping records, quality certificates, pricing agreements, and client communications across multiple languages. Finding specific information required staff to know exactly where to look and often involved manually searching through thousands of records.
We implemented a hybrid retrieval system using both RAG (Retrieval-Augmented Generation) and CAG (Cache-Augmented Generation). Frequently accessed reference data, current pricing, active contracts, supplier specifications, product catalogs, is preloaded into a Redis cache for sub-second CAG responses. Complex historical queries use RAG to search across the full document corpus.
The natural language interface supports queries in Chinese, Russian, and English, reflecting the company's international operations. A sales manager can ask 'What was our average sunflower oil price from Ukrainian suppliers in Q3 2023?' or 'Which Russian partners have increased their volume year-over-year?' and get instant answers with source citations.
The system indexes not just structured CRM data but also unstructured content: email threads, WeChat conversations, scanned contracts, and shipping documentation. Staff can query across all information sources without knowing which system holds the data they need.
Agricultural commodity trading involves repetitive data entry: creating deal records, onboarding new suppliers, updating client information, tracking shipments. Each task follows specific workflows but requires pulling information from multiple sources and entering it into the CRM correctly.
We built specialized agents for the most common workflows, each equipped with the specific tools needed for that task. The Deal Creation Agent can create a new deal record from a conversation summary, it extracts commodity type, volume, pricing, delivery terms, and counterparty details, validates them against existing records, and populates the CRM with minimal human input.
The Client Onboarding Agent handles new partner registration. When Forward East starts working with a new supplier from, say, Kazakhstan, the agent can process their documentation, verify company registration details, set up the appropriate payment terms based on country and commodity type, and create the CRM records. What previously took hours of manual data entry now happens in minutes.
The Supplier Management Agent monitors ongoing relationships: tracking delivery performance, flagging when contracts are due for renewal, and alerting staff when a supplier's quality metrics fall below threshold. Each agent has access only to the tools and data relevant to its specific function, following the principle of least privilege.
International agricultural exports require extensive documentation: bills of lading, phytosanitary certificates, quality inspection reports, certificates of origin, customs declarations, and commercial invoices. Each shipment can involve 10-20 documents, and a single inconsistency, a mismatched weight, wrong HS code, or incorrect date, can delay customs clearance for days or weeks.
Human errors in documentation are common and costly. A quality certificate might show 42% protein content for sunflower meal while the commercial invoice lists 41%. The certificate of origin might have a different shipper name than the bill of lading. These inconsistencies often aren't caught until the shipment reaches port, causing expensive delays.
We deployed QwenVL 3, Alibaba's state-of-the-art vision-language model, to automatically verify export documentation. The system ingests all documents for a shipment, extracts text and data using OCR, and cross-references every piece of information that appears in multiple documents. Weights, dates, commodity specifications, company names, vessel information, and container numbers are all checked for consistency.
When the system detects a discrepancy, it flags the specific issue with visual highlights showing exactly where the inconsistency appears in each document. Staff can review and correct errors before documents are submitted, rather than discovering problems at customs. The system also checks against known requirements for different destination countries and commodity types, catching compliance issues early.
The QwenVL 3 model handles documents in Chinese, Russian, and English, as well as mixed-language documents common in international trade. It processes handwritten annotations, stamps, and signatures that traditional OCR systems struggle with. Processing time is under 30 seconds per document set, enabling verification of every shipment rather than spot-checking.
The intelligent query system now handles hundreds of daily searches across Forward East's historical data. Questions that previously required asking senior staff or digging through archives are answered instantly. New employees can access institutional knowledge immediately rather than building it over years.
The automation agents have reduced manual data entry by approximately 90% for covered workflows. Deal creation that took 20-30 minutes now takes 2-3 minutes of review time. Staff focus on relationship building and strategic decisions rather than administrative tasks.
The document verification system has caught numerous inconsistencies that would have caused customs delays. More importantly, it has changed behavior: knowing that documents will be automatically checked, staff and partners are more careful with accuracy upfront. The error rate in submitted documentation has dropped significantly since deployment.
Forward East can now scale their operations without proportionally scaling their administrative staff. With 500,000+ tons of annual imports and hundreds of active partner relationships, the AI systems handle complexity that would be unmanageable manually.
If your team is one or two unknowns away from a system like this one, a thirty-minute call is the fastest way to find out.
Book a discovery callEngagements range from two-week diagnostics to multi-month builds, scoped after a single discovery call.
Every project on this page shipped because we said no to the wrong scope before we said yes to the right one. Half the value of working with us is the engagement we will not take. The other half is the system that ends up running in your business.
Healthcare, defense-adjacent, and enterprise clients sign NDAs that prevent naming. Engagement scope, technology stack, and measured outcomes can be shared publicly. Client identity stays protected.