The Tension: High Stakes, Broken Foundation

To close major contracts, we needed to integrate complex features like bulk analysis and custom reporting. However, the foundation was crumbling under two specific pressures:

1. The "Artistic" Debt

The previous designer created a system that was visually stunning but structurally rigid. It was built for a "happy path" gallery.

• The Component Trap: Instead of a scalable modal with variant properties, we had a new variant for every single scenario.

• Orphan Screens: The designs were static snapshots. There were no user flows, no error states, and no logic for when the LLM failed.

2. The Communication Vacuum

Sales kept promising features that Engineering couldn't estimate because the specs were vague. The CTO was burning out trying to act as Product Manager, Architect, and Designer all at once.

My Role: The Translator & Architect

I didn't just push pixels; I functioned as a strategic mediator. I used my technical background to sit with Sales and translate their "Why" into technical user stories that the CTO could actually build.

• The Tactic: I used the Weekly Decision Meeting to facilitate real-time trade-offs. If Sales wanted a new feature, I explained the technical debt it would create. We moved from a "Wish List" to a realistic Roadmap.

The Craft: Designing for AI Complexity

My core design challenge was moving the user from "playing with a chatbot" to "managing an enterprise workflow." I introduced four pivot features that transformed the UX:

1. Scaling RAG: From Single File to Collections

• The Context: The engineering team achieved a massive breakthrough: enabling RAG (Retrieval-Augmented Generation) across entire folders of documents, not just single files.

• The Design: The temptation was to build a complex new interface. Instead, I chose invisibility. I introduced a simple "Scope Switch" tab (Docs vs. Collections).

• The Win: By using a familiar UI pattern to trigger a complex backend capability, we kept the learning curve at zero while expanding the tool’s power exponentially.

  

2. Async Multi-Tasking

• The Problem: The UI blocked the user while the AI "thought" (sometimes for 30+ seconds).

• The Solution: I introduced a background processing state. Users could now upload files and start new analyses while previous ones processed.

• Impact: This friction removal contributed to a 50% increase in Session Duration.

3. Chained Logic Workflows

• The Feature: We built a builder that allowed users to launch specific prompts based on the results of a previous prompt.

• The Craft: Visualizing invisible logic. I designed a node-based view where users could see how Prompt A (Extract Date) fed into Prompt B (Verify Deadline).

4. User-Defined Conditional Logic

• The Feature: Users needed to scan massive reports for specific risks.

• The Craft: I designed a "Keyword Highlighter" system. Instead of relying on AI to guess what mattered, users could define their own rules—custom color-coding for specific answers, or filtering documents that started with specific project codes (e.g., "HS2-").

• Impact: This shifted the UI from passive reading to active auditing, significantly speeding up data absorption.

The Real-World Test: HS2

All these features came together for the HS2 (High Speed 2) project. They needed to analyze thousands of "Verification and Validation Certificates" (VVCs). By using the Collection feature to ingest files and the Chained Workflows to cross-reference dates, we turned a manual nightmare into an automated flow.

The BuildPrompt team collect their trophy from Karolina Orecchini of Digital Construction Week (far left), Lilian Ho of Aecom (second right) and Sally Phillips (far right) (Photo © 2025 – ASV Photography Ltd. www.ASVphotos.com)

Delivery: Building with Handcuffs On

We didn't have time to rewrite the codebase. I had to be pragmatic. I kept the legacy design system and applied tweaks to design the new features. I often shared a "North Star" vision, a mockup of how it could look to keep the team inspired while we shipped the MVP.