Projects

Project Overview

Replicating expert research methodologies with intelligent agents to open the AI black box, enabling enterprises to rapidly generate explainable and traceable business analysis reports.

Industry Pain Points

Insufficient Control in Research Process

Industry research informs critical business decisions. Experts demand controllable problem decomposition and data collection methods, which current AI assistants cannot provide with sufficient transparency and control.

Delayed Response to Market Changes

• Rapid updates in automotive technology and consumer preferences
• Market departments in large automotive companies struggle to capture consumer demand changes in time

Limitations of Traditional Research Methods

• Subjectivity in manual market research and analysis
• Limited information channel breadth in manual research

Innovative Solution

Connecting intelligent agents to search engines, internal enterprise knowledge bases, and targeted resources. Combined with the agent's own tool orchestration capabilities, the system generates data-driven business insight reports and significantly improves research efficiency.

Core Methodology

Using large language models and AI search engines, based on client market research methodologies, replicating the entire analysis workflow while incorporating human supervision and discussion stages to control information reliability while expanding information channel breadth.

Technical Implementation

• Using search agents and citation pool injection, filtering, and maintenance to ensure information validity
• Context management
• Traceability for each analysis stage
• Human-in-the-loop mechanism to collect user preferences and build enterprise knowledge
• Reasoning with extensive information
• Consumer demand scenario simulation
• Generating forward-looking reports and visualized analysis results

Application Scenarios

Research including automotive industry consumer demand, in-vehicle experience competitive analysis, consumer behavior research, etc.:

Through AI replication of expert analysis frameworks, automatically decomposing analysis topics, generating complete analysis reports including social trends, market size, consumer characteristics, user personas, etc.

Project Overview

An AI-powered smart customs system enabling end-to-end intelligent processing from trade documents to customs declarations, transforming traditional manual workflows into efficient, data-driven operations.

Industry Pain Points

Low Efficiency & High Error Rate

Customs declaration involves processing a massive volume of marine trade documents. Manual data entry is slow and prone to errors, leading to potential delays and compliance risks.

Complex Decision Making

Accurately identifying HS codes and calculating taxes requires significant expertise. Manual lookups are inefficient, and calculation errors are common in complex scenarios.

Knowledge Retention Challenges

Customs clearance experience is often tacit knowledge held by individuals. It is difficult for enterprises to accumulate, digitize, and reuse this historical data effectively.

Innovative Solution

Core Methodology

The system leverages AI to automate the entire workflow:

1. Intelligent Document Processing: Using AI OCR to recognize multiple document types (Invoices, Packing Lists, Bills of Lading, Arrival Notices) and automatically cross-checking consistency across different files.
2. Smart Decision Support: Implementing intelligent HS code matching and tax rate judgment. The system automatically calculates critical metrics like duties and volume, while accumulating historical declaration data by client to refine future predictions.
3. Human-AI Collaboration: "AI Preparation + Human Review" model ensures accuracy while seamlessly integrating with customs declaration systems in China, Singapore, and Japan.

Technical Implementation

• Hybrid Retrieval Engine: Utilizing Qdrant for vector search combined with traditional keyword search to accurately retrieve HS codes and regulatory information.
• Multi-Modal Data Processing: detailed extraction and structuring of data from unstructured trade documents.
• Knowledge Graph: Building a domain-specific knowledge base for customs regulations and historical cases.

Value Delivered

• Efficiency Surge: Reduced manual processing time from 1-2 hours to minutes.
• Error Reduction: Significantly lowered the error rate in data entry and calculations through automated cross-validation.
• Digital Inheritance: Successfully achieved the digital accumulation and inheritance of customs clearance experience, reducing dependency on individual experts.

Project Overview

A one-stop pharmaceutical distribution data governance platform built on a big data architecture. It addresses the challenges of data fragmentation and analysis latency in the China region, enabling real-time monitoring and intelligent analysis of sales data for large-scale pharmaceutical enterprises.

Industry Pain Points

Data Fragmentation & Integration Difficulty

Sales data is scattered across various distributors and channels with inconsistent formats. Manual aggregation is inefficient and prone to errors, making it difficult to form a unified view of the market.

Analysis Latency

Traditional systems struggle to process tens of millions of sales records efficiently, leading to significant delays in business reporting and missed decision-making windows.

Lack of Standardization

The absence of unified master data standards for drugs, institutions, and personnel hinders effective data asset management and cross-dimensional analysis.

Innovative Solution

Core Methodology

The system establishes a closed-loop data governance framework:

1. Unified Data Ingestion: Automated collection and cleaning of flow data from multiple sources.
2. Master Data Management (MDM): rigorous standardization of core data dimensions (Products, Hospitals, Representatives).
3. Intelligent Analytics: Real-time calculation of sales metrics with automated appeal handling workflows.

Technical Implementation

• High-Performance Architecture: Adopting Spring Cloud microservices backend combined with Spark for big data processing to ensure scalability.
• Hybrid Storage Layer: Utilizing PostgreSQL for relational data, Elasticsearch for high-speed search, and Redis for caching.
• Enterprise-Grade Security: Implementing Shiro + JWT for robust access control and data security.
• Observability: Integrated Prometheus + Grafana for real-time system monitoring.

Value Delivered

• Massive Data Processing: Successfully achieved real-time processing capability for over 10 million flow data records.
• Efficiency Leap: Through digital transformation, operational costs were reduced, and overall management efficiency improved by more than 40%.
• Standardization: Established a unified data standard system, transforming raw data into valuable corporate assets.

Project Overview

An intelligent literature analysis system designed for Pharmacovigilance (PV). It automates the screening of medical literature to identify company products and assess safety information, ensuring compliance and patient safety while significantly reducing manual workload.

Industry Pain Points

High Volume of Literature

Pharmacovigilance teams must monitor a vast amount of global medical literature daily. Manual screening is labor-intensive and inefficient.

Risk of Missed Safety Signals

Identifying adverse drug reactions (ADRs) and special situations (e.g., overdose, pregnancy use) requires high precision. Manual review runs the risk of missing critical safety signals.

Standardization & Traceability

The regulatory requirement for PV is strict. Manual processes often lack consistent scoring standards and traceable audit trails for decision-making.

Innovative Solution

Core Methodology

The system employs a multi-stage AI pipeline to process literature:

1. Automated Retrieval & Matching: Automatically retrieves literature meta-data via APIs and matches them against the company's product catalog.
2. AI Relevance Scoring: The system downloads PDFs and uses NLP to analyze content, learning from historical data to calculate a relevance score (0-100%) for each article regarding company products.
3. Safety Signal Detection: Automatically identifies "Safety Information" (e.g., side effects).

Identified: Auto-tag as "Contains Safety Info". Clean: Auto-tag as "No Safety Info".

* Uncertain: Triggers a "Human-in-the-loop" workflow, sending email alerts with the document ID and reasoning to experts for manual review.

Technical Implementation

• AI-Driven Analysis: Utilizing Large Language Models (LLMs) to perform semantic analysis on full-text PDFs.
• Active Learning: The scoring engine iteratively improves by learning from 100% relevant historical cases, capturing context patterns and keywords.
• Intelligent Assistant: Integrated AI Agent + MCP (Model Context Protocol) to enable natural language semantic search and Q&A across the literature database.

Value Delivered

• Efficiency: Automates the initial screening of thousands of documents, allowing experts to focus only on high-risk or uncertain cases.
• Compliance: Ensures no potential safety signal is overlooked through rigorous AI scanning and automated alerting.
• Agility: The semantic search capability allows teams to quickly retrieve specific safety data or historical cases using natural language.