Infraclear: Designing Trust Into Infrastructure Data
An AI-native infrastructure intelligence platform helping professionals turn 200,000+ dense deal documents into answers they can verify, act on, and defend.
Built on 222,000 agreements and 3.7 million extracted terms, Infraclear serves contractors, lawyers, project financiers, and investment teams who make multi-million-dollar decisions from documents that run hundreds of pages.
My Role
Founding Product Designer
Information Architecture
UX/Product strategy
Duration
4 Months
(Jul 2025 - Oct 2025)
Collaborators
Founder & Co-founder
Data Scientists
Front-End Engineers
The Problem
Infraclear’s dashboard buried deal terms in nested views. Comparing two deals meant going into one, backing out, then opening the other, and losing context each time. AI added a harder problem: lawyers and financiers couldn’t trust an answer they couldn’t trace.
The Solution
The redesign gave users one view to compare deals without losing their place. One navigation model now runs across every surface, so the platform works as a single system. Every AI answer links back to its source, so it’s fast and easy to verify.
What difference did we make?
The Solution
Two connected pieces of work, designed as one system. A restructured dashboard that keeps users oriented while comparing deals, and an AI chatbot built on progressive disclosure that answers natural-language questions in the data-grid format its users already trust.
Oriented, Not Lost
Persistent filters, contextual breadcrumbs, and inline comparison modules let users evaluate multiple deals without losing their place or retracing steps.
Answers You Can Verify
Every chatbot response is anchored to a source document and expandable into a structured data table, so users can check the work rather than take it on faith.
Familiar by Design
The answer view is built around a data grid, not a wall of text, matching the Excel-based mental model contractors and lawyers already bring to the job.
Process
Mapping Every Path, Not Just the Happy One
I started by mapping all user pathways and edge cases across the platform, then worked with PMs and domain experts to identify where people were actually getting stuck. Every interactive state got documented: account menus, filter dropdowns, saved filter modals, collection panels, empty states.
Annotated dashboard audit
The Friction: Task Switching
The pattern that kept surfacing was task switching. Comparing financial terms across two or three deals meant navigating multiple nested views, then navigating back, losing context each time. In a job that already demands intense attention to detail, that constant re-orientation was quietly draining focus.
The Structural Fix
Persistent filters — that stay active across views, so context never resets.
Contextual breadcrumbs — that always answer where am I and how do I get back.
Inline comparison modules — that let users compare terms without leaving the page.
Applied consistently across Agreements, Datafinder, Models, and Find & Compare Terms, so the platform speaks one navigational language everywhere.
Agreements, Datafinder and Models views
Impact
[X]%
reduction in clicks to complete a cross-deal comparison
[X]%
improvement in time-on-task
Users reported feeling more oriented and able to complete workflows without retracing their steps.
Chapter Two: The AI Chatbot
The Challenge
The chatbot wasn’t a feature bolted onto the dashboard. It was meant to become the anchor for the entire platform, the first thing users touched and the lens through which everything else made sense. The range of possible questions was wide and loosely defined, and the experience had to feel systematic on the first try, not something users had to learn.
Structuring Around Progressive Disclosure
I broke the experience into three stages so users could start broad and narrow in gradually, instead of facing a blank chat box.
01
Welcome View
Entry points organized into clear, topic-based buckets. Instead of a generic search bar, users land somewhere that tells them where to start. Findability and wayfinding, applied to a conversational interface.
02
Intermediate View
Context-aware follow-up prompts guide the conversation forward, so users don’t need to know exactly what to ask next.
03
Answer View
Anchored in a data grid rather than generated prose. Contractors, lawyers, and infrastructure professionals already treat Excel as their source of truth, so the answer format meets them there. A dataset selector lets users control which sources the model draws from, adding flexibility without adding clutter.
Chatbot welcome, question and answer views
Designing for Enterprise Constraints
Moving toward a live engagement with a sovereign fund’s infrastructure department, delivered through EY, meant the design had to hold up against real requirements:
Data sovereignty — in-region hosting on AWS Saudi Arabia to comply with local data regulations, which shaped what could run where.
Role-based access control — different user groups needed different visibility into sensitive deal data.
Verifiable answers — every response traceable to a source document, with the system trained to recognize the edges of its own knowledge rather than guess.
Concrete use cases — the proposal was grounded in scenarios the client’s team would recognize immediately, like water and electricity distribution projects, rather than pitched in the abstract.
Translating a room full of stakeholder requirements into a scoped design proposal, a set of use cases, and a delivery roadmap was as much a part of this work as the interface itself.
The System Behind the Interface
95%
extraction accuracy
100K+
documents processed monthly
90%
lower compute cost vs LLM-only
12,000
expert-labeled agreements
Design decisions grounded in a system that could actually perform at enterprise scale is what made the proposal credible rather than aspirational.
Closing the Loop
QA That Protects Trust
Shipping v1 wasn’t the finish line. Once real usage started, I tracked and specified a set of interaction bugs that were quietly eroding trust:
FIXED
Feedback controls were scrolling users back to the top of a response, an unnecessary interruption that got cut.
FIXED
Sticky navigation now persists for the full height of the active answer, then hands off to the next question on intersection, instead of jumping early.
FIXED
Selecting a recommended follow-up question now updates the active index automatically, matching behaviour already built into the welcome screen.
FIXED
Data table state bugs resolved: results disappearing mid-scroll and showing No results found until a user toggled views, and in one case a question’s table displaying results belonging to a different question.
Specifying these precisely, with reproduction steps rather than “the table is broken,” was part of protecting user trust in a product whose entire premise is giving people reliable answers.
What’s Next
NEXT
A feedback modal on negative ratings, asking users to categorise why an answer missed (inaccurate, wrong sources, no information), turning every piece of criticism into structured data for improving the model.
NEXT
Autocomplete in the welcome view, helping users complete prompts as they type rather than starting from a blank field.
Reflections
Trust Is a Constraint, Not a Feature
In an AI product, every answer needs a path back to its source. That requirement shapes every other design decision downstream.
Systems Thinking Beats Redesign
The dashboard didn’t need rebuilding. It needed one seam fixed, the point where friction was compounding across every workflow.
Following Through Is the Job
Shipping v1 is the easy half. Catching and specifying the small interaction bugs is what keeps a product trustworthy after launch.
