Exemble Designing an LLM workflow environment for internal knowledge
Exemble is an on-premise AI platform where teams connect agents powered by internal data and proprietary models into workflows, then complete a variety of tasks through chat.
I designed the core experience across user chat and workflow building as well as administration, permissions, and monitoring. As the team's sole designer, I led the project from early service planning onward.
Chat workflows powered by internal data
Unlike a general-purpose LLM chat product, Exemble must manage internal data, external APIs, models, and permissions together. Users can generate answers from company documents or run workflows to produce work outputs, while administrators need to operate and govern the entire process.
Separating user and administrator experiences
The product is divided into a user space for chat and workflow execution and an administrator space for managing the resources behind them. I varied the layouts and information density so the user interface supports a fluid task flow while the administrator interface supports complex resource management.
Building complex work directly with workflows
The workflow builder lets users combine agents, API tools, and conditional branches into real business processes. Nodes and edges visualize the flow, but the result is more than a diagram: it is an executable structure with status feedback, testing, and debugging.
A workspace that preserves the user's train of thought
Relationships and context between nodes are essential when designing work. Keeping every node editable in one workspace preserves that context, while shortcuts, automatic layout, and draft saving make continuous iteration and validation feel natural.
Trustworthy chat from question to result
Chat goes beyond question and answer: it produces results and executes tasks according to the selected workflow. Showing the process step by step helps users trust the system's decisions and quickly locate the cause when a result needs improvement.
Evidence users can inspect directly
Sources used for an answer are shown alongside the result so users can inspect its origin and context. The evidence can come from attached files or previously indexed internal knowledge, creating differentiated value and increasing confidence in the response.
Connecting varied internal data with less effort
Agent knowledge is collected in different forms through database connections and document uploads. Context-specific browsing patterns and automatic re-vectorization when data changes help teams build and maintain a useful knowledge base.
Model operations connected to performance improvement
Chat models are managed through status, resource usage, and change history. Integrating fine-tuning information for each model creates an operational environment that supports continuous performance improvement rather than simple service maintenance.
A hierarchical operations dashboard
The dashboard moves progressively from all workflows to an individual workflow and its resources, with color communicating status. Administrators can notice operational issues quickly and follow a natural path toward the underlying cause.
Making an AI service feel alive
Motion appears throughout workflow connections, answer-generation feedback, and chat-session transitions. These interactions are not merely decorative: they bring the AI experience to life and make a complex enterprise system easier to understand.
Leading a project in an engineering-led team
Exemble turned complex, abstract LLM concepts into a working product in a short period. In an all-engineering team, I collaborated closely on technical constraints and implementation structure while applying a design system for consistency and future expansion.
Working as both planner and designer, I contributed to the product direction and design strategy and gained experience leading the project end to end.