Unified Chatbot Framework at Jeppesen (Boeing)

The Problem

Jeppesen’s enterprise AI engineering team faced a growing challenge: internal teams across the organization were independently piloting and experimenting with AI-powered chatbots. This fragmented approach led to duplicated work, compliance hurdles, and isolated deployments with no shared learning or reusable components.

The Manual Process:

• Each team building chatbots from scratch independently
• 512 hours to build a single chatbot or agent from scratch
• Every team navigating AI governance and IT policy approvals separately
• No shared templates, components, or knowledge base
• Custom interfaces, deployment models, and tool integrations per team

Key Pain Points:

• Redundant effort across multiple teams building similar components
• Compliance burdens requiring each team to individually handle governance
• Long development cycles preventing rapid experimentation
• Lack of standardization creating maintenance nightmares
• No way to share learnings or improvements across teams

The Solution

A lean 5-7 person frameworks team built the Unified Chatbot Framework (UCF)—a templatized platform that dramatically simplifies agent creation using LlamaIndex’s open-source components. The framework provides shared infrastructure with built-in security, compliance, and privileged access to internal systems.

Impact: Reduced development time from 512 hours to 64 hours per agent (87% reduction). Already saved 1,792 collective hours across teams, with projections of ~4,900 hours saved annually after global rollout. Now powering 10-11 production products across Boeing.

How It Works

Key Technologies:

• LlamaIndex Event-Driven Workflows - Flexible agent orchestration with session and state management
• Tool Connectors - GraphQL, SQL, REST APIs, Neo4j, Databricks Data Mesh
• Multi-LLM Support - Azure, AWS, HuggingFace, open models
• Vector Database Options - Azure AI Search, Qdrant
• Conversation Memory - Azure CosmosDB, Azure Table, Azure Cache for Redis

Process Flow:

1. Teams start with pre-built workflow templates for single or multi-agent systems
2. Import a template and configure with a simple JSON config file
3. Bring your own LLM, vector database, and data sources
4. Connect to internal systems using pre-built tool connectors
5. Build custom tools based on your data without worrying about base LLM app components
6. Deploy with built-in enterprise compliance and security standards
7. Framework handles governance, privileged access, and standardization automatically

Developer Experience: Agents can be built with ~50 lines of code and a JSON config file, with all enterprise requirements handled automatically.

Key Insight

Jeppesen evaluated several agent frameworks, including chain and graph-based options, but chose LlamaIndex for its flexibility and production-grade control. Rigid graph-based frameworks couldn’t meet enterprise constraints.

Why This Matters:

• Event-driven workflows with full control over state transitions vs. rigid predefined flows
• Transparent, open-source tooling with strong developer community support
• Fine-grained control over agent logic, inputs, and execution flow
• Modular design allows bringing your own components (LLM, vector DB, data sources)
• Low-code configuration (JSON file) while maintaining full customization capability
• Built-in support for custom tool creation and template injection

Platform Evolution: UCF started as a chatbot framework but is naturally evolving into a full agent orchestration system. Because LlamaIndex treats everything as a workflow, it supports both conversational and automated event-triggered tasks.

Success Measurement: “We measure success not just in hours saved, but in how quickly teams can explore, contribute back, and scale securely.” —Karan Dodiya, Technical Product Owner

Links

• Original Blog Post - Full case study
• LlamaIndex - Agent framework used
• LlamaIndex Workflows - Documentation on event-driven workflows