I spearheaded the development of the company's first AI-powered component using Google Gemini Model. This chatbot helps construction contractors quickly extract key information from 200+ page project documents, reducing project assessment hours to minutes. The MVP design improved bidding efficiency by enabling contractors to evaluate more projects in less time while reducing the risk of overlooking critical requirements that could lead to costly change orders.
Lead Product Designer for our Team of 10. Worked closely as a "triad" with Designer+PM+Lead Engineer along with 8 backend and front-end engineers, and a ML architect.
Construction contractors spend hours manually reviewing 200+ page document sets to extract key project information for bidding decisions. This time-intensive process creates bottlenecks and often leads to missed opportunities or costly oversights. The UX Callouts below illustrate potential questions our users might ask.
Our AI-powered solution is integrated into the Project Detail Pages for workflow integration. By allowing users to ask questions in natural language, the chatbot provides fast, accurate responses that are directly sourced from the projects documents. This not only saves time but also enhances decision-making by ensuring that contractors have access to the most relevant information without the need for manual document review.
“This would save me hours every week. I can finally get answers without reading through hundreds of pages.” - General Contractor, Dallas
We conducted weekly user interviews throughout the design process, testing assumptions and validating design decisions in real-time. This allowed us to quickly iterate on features like suggested prompts and trade filtering, pivoting from a modal to a drawer interface after observing users' need to reference documents while chatting.
By adopting a build-measure-learn cycle with weekly sprints, we were on the road to delivering a fully functional MVP within two months. This approach allowed us to address the core user need efficiently, transitioning from problem identification to a live product much faster than traditional processes.
Through research, I identified three user personas: primary users are general contractors, secondary are subcontractors, and tertiary are estimators. I mapped the contractor bidding journey, pinpointing pain points where an AI assistant could add value, revealing contractors' need for varied information at different stages.
“How do addenda changes affect my prepared bid?”
“What exactly is required for my trade?”
“Is this project one that's worth us pursuing now?”
The user journey map was instrumental in shaping both the interface design and the AI training requirements. It also helped validate task flows with stakeholders before committing to high-fidelity designs. This map outlines the journey of our primary personas, General Contractors, as they search for new projects to bid on and review necessary documents before submitting their bids.
We believe AI will streamline this process significantly.
Through extensive research and my understanding of assumptions mapping, inspired by Teresa Torres' Continuous Discovery Habits, I developed a method to visualize the narrative and assumptions. I included a section for the AI Assistant to illustrate the behind-the-scenes processes, highlighting the backend work and when it occurs within the whole story. Here's an example of one of the many maps like this I created:
I compiled a comprehensive list of our MVP components and documented the core AI capabilities for team reference. I participated in complex architecture and backend discussions, which also required visual aids unbeknownst to others. I was involved in high-level discussions about technical feasibility from the outset. Additionally, I considered testing and quality engineering requirements, as I collaborate closely with engineers and QE professionals on a daily basis.
I needed to create something that aligned with our current brand guidelines and adhered to the existing design library. I aimed to avoid introducing entirely new components, as our users typically do not respond well to unfamiliar elements. With this in mind, I successfully adapted our existing components to align with the new integrated approach, ensuring a seamless introduction of AI within our highest-valued product offering.
Early testing also helped inform a cut down and cleaner style guide to reduce visual overload.
I prefer to begin with traditional pen and paper sketches. These initial drawings help me brainstorm multiple iterations and approaches before transitioning to low-fidelity versions. Once I have two or three strong concepts, I create task flows to narrate their stories and document UX callouts.
We explored the option of displaying the AI assistant within a modal. It provides a focused interaction space and a large area for conversation. But, this idea obstructed access to source documents during chats for reference. Not ideal.
We identified that when we maintained the PDP and document context, it provided reference while chatting. When users could see all the context at once, it made it easier for them to use our new LLM as a result.
The colors needed to stand out from the rest of the project details page and exude the fun and playful tone that we see widely used in other products AI branding. Testers praised the page's colors as 'bright, playful, and fun,' with a strong preference for purple.
The final design decisions included placing the AI integration LLM in a right-sided drawer with toggle functionality. The primary interaction involved suggested prompts, with a fallback to free-form text input. For error handling, we implemented graceful degradation with retry options and clear error messaging. We also ensured high contrast ratios to accommodate construction site lighting conditions for better accessibility.
Domain expertise is crucial! Gaining a deep understanding of construction workflows was vital for making informed design decisions. Additionally, acquiring technical knowledge about LLMs demystified their functionality and operation helped me out a LOT.
Examining our iterations enlightened our need for modifying the language and display of information. I came up with several different "UI Trains" to work towards after reaching our MVP. Here's where we were going!
There is always room for growth, and I am truly impressed by how much the team achieved in just two months of dedicated effort. Some of our long-term goals rely on our AI/ML team developing algorithms for us. Here's the direction we planned to take, dependent on survey data and user tests, of course!
