An AI-assisted workspace for Rocket Close that combines automation and human judgment — completing what it can, and guiding teams through complex workflows with clarity, structure, and confidence.
2026
Sr. UX Designer · 0→1 Lead
Rocket Close & Rocket Mortgage
Internal web workspace
Rocket Close is a subsidiary of Rocket Companies that manages the title services and closing of the mortgage process.
UX leadership (Brandon, Director); partner designer (Lauren); and 3 months of daily cross-functional war rooms with architects, developers, analysts, and partner teams. I owned end-to-end UX from 0→1 concept through high-fidelity prototypes and validation.
Nova AI is an internal, AI-assisted workspace designed to support complex, judgment-heavy workflows across Rocket Close. It follows a dual-path model: fully automating tasks across systems when possible, and introducing a human only when judgment, exceptions, or time constraints require it. At that moment, Nova shifts from operator to guide — replacing fragmented systems and manual processes with a single, guided experience that brings together data from Atlas, Salesforce, and vendor tools into one place. It summarizes what's been done, surfaces key context, and presents clear next steps so decisions can be made quickly and confidently.
Result: high cognitive load, slower throughput, and workflows that didn't scale — especially in scenarios involving ambiguity and risk.
Scheduling and assignment are high-volume, time-sensitive workflows — but they rely on fragmented systems, manual coordination, and undocumented decision-making.
Team members were forced to gather context, track outreach, and make judgment calls across multiple tools, often under tight deadlines. This led to inefficiencies, inconsistent outcomes, and unnecessary cognitive load.
Lead UX Designer — 0→1 concept → execution. Nova AI began as a 0→1 concept co-created with UX leadership. I now lead the vision, strategy, and design execution.
I owned end-to-end UX: defining the product direction, establishing human–AI interaction principles, and designing the system across research, workflows, conversation design, and prototyping. In the early stages, there was no PM or PO — I partnered directly with leadership to shape the product, then drove it forward through high-fidelity prototypes to align teams and secure executive buy-in.
Reduce time and effort required to complete scheduling and assignment tasks — while increasing consistency, visibility, and confidence in decision-making.
Design a system that could automate where possible, guide where needed, and scale across workflows without compromising accuracy or operational constraints. Surface only what matters, make reasoning transparent, and preserve human judgment in high-stakes workflows.
Instead of asking "How do we help users read documents faster?" we asked: what if they didn't have to read them at all?
Scheduling notaries and appraisers is a high-frequency, high-variability workflow involving vendor availability, pricing negotiation, time constraints, client dependencies, and exception handling. Today it requires calling multiple vendors, tracking responses manually, juggling tradeoffs in real time, and switching between Atlas, email, notes, and internal tools.
Team members spend significant time coordinating outreach, tracking vendor responses, and evaluating tradeoffs (fee vs. time vs. experience). The process is inconsistent, mentally taxing, difficult to scale, and highly dependent on individual judgment.
Instead of asking "How do we help users schedule faster?" we asked "How do we structure decision-making so users don't have to manage complexity themselves?"
In a common scenario, a team member must identify the correct individual from a judgment or public record document containing hundreds of repeated names — for example, a document referencing "John Smith" dozens of times across many pages. Today this means scanning page by page, comparing attributes (birthdate, SSN fragments, addresses, middle initials), holding partial matches in memory, and spending hours verifying identity. It's cognitively expensive, error-prone, and poorly suited to human pattern-matching at scale.
Team members were acting as human parsers of unstructured data — manually scanning large documents, comparing fragmented attributes, and holding partial matches in memory before making a decision. In identity matching workflows, this could take 30–90 minutes per file, with high cognitive load and risk of error.
Through shadowing and journey mapping, we found the core problem wasn't speed — it was who does the first pass of reasoning. Humans were doing work that machines are fundamentally better at: pattern matching at scale, cross-referencing large datasets, and filtering ambiguity.
Nova AI reverses the workflow. Instead of human investigates → then decides, we designed AI investigates → human decides. Nova parses entire documents instantly, identifies candidate matches, cross-references known attributes, and surfaces only decision-ready insights. Each result includes confidence level, matching attributes, missing data, and plain-language reasoning.
Team members spend time gathering and verifying information before they can act.
Nova performs investigation upfront — aggregating data, checking sources, and surfacing only what's needed before the task reaches the user.
Users must figure out what to do next in complex, ambiguous workflows.
Tasks are framed as clear decisions, guiding users step-by-step instead of requiring them to manage the workflow themselves.
Work is fragmented across multiple systems (Atlas, Salesforce, email, etc.), requiring constant context switching.
Nova consolidates tasks, context, and actions into a single workspace — reducing system hopping and cognitive load.
Critical information is buried across tools, requiring users to search, scroll, and remember details.
Nova surfaces a structured, decision-ready summary — key details, requirements, and prior actions in one place. Users wanted important info "upfront without scrolling."
Users lack key information when selecting vendors, leading to delays or poor decisions.
Vendor cards provide relevant details — availability, requirements, history, and (future state) performance — to enable faster, more confident selection.
Handling changes (fee, experience, time) is inconsistent and requires manual judgment each time.
Nova standardizes change handling with guided flows — different logic paths for fee, experience, and time/due date — helping users evaluate and act with clarity.
Users must manually determine when approvals are needed, slowing down workflows.
Nova evaluates thresholds automatically — auto-accepting small changes, routing larger ones for approval or fallback. Especially strong for notaries (~2 min approvals); appraiser approvals (hours-long) are an automation opportunity.
When ideal options fail, users must remember and re-evaluate all previous attempts manually.
Nova saves and ranks fallback options by impact, allowing users to quickly choose the best alternative. The holding queue concept was understood and well received in testing.
Users don't trust AI when they can't understand how decisions are made.
Nova explains its recommendations — showing reasoning, confidence, what was checked, and what's uncertain. From testing: "Trust depends on completeness of information."
Manual outreach and coordination — calls, emails, follow-ups — consume time and are hard to track.
Nova attempts outreach, logs actions, and adapts next steps. It steps in with human support only when needed.
Approvals (especially fees) are slow and disconnected from the workflow; data must be pulled from multiple systems to complete tasks.
Nova initiates and tracks approvals directly within the flow, and integrates data from Atlas, Salesforce, and vendor systems into a single, unified view.
Shadowed notary and appraisal schedulers across Nova, Atlas, and Salesforce; mapped decision workflows and edge cases. Explored early concepts through "vibe coded" prototypes to validate AI-led investigation.
Identified core friction: decision-making was fragmented across systems. Reframed workflows into structured decision models and defined patterns for AI reasoning, thresholds, and fallback logic.
Built high-fidelity Figma prototypes simulating AI behavior — scheduling flows, change handling, and ranked fallback. Iterated rapidly with another UX designer (Lauren) across notary and appraisal use cases.
Tested with notary and appraisal schedulers through guided prototype sessions. Ran 3 months of daily cross-functional "war rooms" (UX, engineering, architecture) to validate feasibility, edge cases, and scalability.
We conducted shadowing sessions with notary and appraisal schedulers, mapping how decisions were made across Atlas, Nova Workspaces, Salesforce, email, and vendor tools. A clear pattern emerged: team members were acting as coordinators of fragmented systems, manually gathering information before they could make a decision.
In parallel, we explored early concepts through rapid "vibe coded" prototypes — using lightweight builds to simulate AI behavior and validate whether shifting investigation to the system was even viable. These early experiments helped us quickly test the core idea: what if the system did the work before the human? This phase reframed the problem from a workflow inefficiency to a decision-structure problem.
Nova AI began as a 0→1 concept led by UX — initially defined by myself and the UX Director, Brandon — before expanding into a broader cross-functional effort. Using Figma, we created high-fidelity, end-to-end flows that simulated real-world behavior, including AI reasoning, change handling, and fallback scenarios.
As scope expanded, I partnered closely with another UX designer, Lauren, to iterate on scheduling and assignment flows across notary and appraisal workflows. To ensure feasibility and scalability, we ran three months of daily two-hour "war room" sessions with architects, developers, analysts, and partner teams — aligning system behavior, edge case handling, data dependencies, and long-term scalability across teams. The product was shaped collaboratively in real time.
We tested scheduling and assignment flows with two primary groups: notary schedulers and appraisal schedulers. Participants were guided through happy paths and complex edge cases, including vendor changes, threshold scenarios, and fallback decision-making.
Users quickly understood and trusted the guided flow model. The holding queue concept was intuitive and well received. Gaps in information surfaced early and were iteratively resolved. Users consistently expressed excitement and confidence in the system: "This will make my life so much easier." · "I love this." · "This is so cool."
Because Nova AI operates in high-stakes, judgment-heavy workflows, we focused on creating a voice that was clear, direct, and neutral — supportive without being overly prescriptive, and transparent about uncertainty.
Rather than replacing human judgment, Nova AI communicates what it has already done, what it knows, and what it still needs from the user. Conversation design emphasized guided clarity over automation, helping users feel in control while still being supported.
Many workflows involve edge cases, exceptions, and risk that cannot be fully automated. Over-automating could lead to incorrect decisions; under-automating keeps workflows inefficient.
Bounded autonomy model: Nova completes investigation and guides decisions, but keeps the human in control at critical points.Team members rely on multiple systems (Atlas, Salesforce, vendor tools, email) to gather information, leading to context switching and incomplete visibility.
Unified AI workspace: Nova aggregates data, surfaces key context, and enables action without leaving the system.Many workflows weren't formally documented — decision-making lived in team members' heads, shaped by training and edge-case experience. This made it difficult to define consistent logic, identify patterns, or translate behavior into a scalable system.
Collaborative workflow extraction: repeated shadowing, workshops, and 3 months of daily cross-functional "war rooms" to externalize decision logic, map edge cases, and translate tribal knowledge into structured, system-driven workflows.Nova needed to support fundamentally different workflows — scheduling, title examination, document classification — each with unique logic, inputs, and outputs. Bespoke experiences for each wouldn't scale.
Modular system framework: a consistent three-panel workspace (tasks, conversation, actions) adapts across use cases, while components, data, and logic shift based on the workflow.Critical decisions were tied to reading and interpreting long, unstructured documents — slow, cognitively demanding, and difficult to scale.
Decision-first interface: Nova removes the document as the starting point, presenting extracted, structured insights while preserving access to source material when needed.Most real-world scenarios don't follow the "happy path" — vendors request changes, don't respond, or require tradeoffs that are hard to track manually.
Dynamic change handling and a holding queue: exceptions are captured, ranked by impact, and reintroduced as structured fallback decisions.In a regulated, high-stakes environment, users won't rely on AI unless they understand and trust its reasoning.
Transparent reasoning: Nova shows what it checked, why it recommends an option, and where uncertainty exists.Scheduling decisions (fee, time, experience) varied widely depending on individual judgment, training, and experience.
Structured decision framework: thresholds, guided flows, and ranked fallback options standardize how decisions are made.
We designed and prototyped a new AI-assisted scheduling workflow, including vendor recommendations, change handling (fee, time, experience), and a ranked holding queue for exception scenarios. The MVP focused on reducing manual coordination, surfacing decision-ready information, and enabling team members to complete tasks without leaving Nova.
Rather than a traditional release, this work was validated through iterative prototypes, stakeholder reviews, and user testing with scheduling teams. Each iteration focused on clarity of decision-making, reduction of manual steps, and confidence in AI-supported workflows before moving forward.
The most effective use of AI wasn't automation — it was structuring decisions. By framing tasks as clear questions with bounded actions, Nova helped users move faster without removing their judgment.
Users didn't need more tools — they needed the right information at the right moment. When we surfaced key details (availability, constraints, impact), users could make decisions confidently without leaving the system.
The "happy path" is rarely the reality. Designing for changes, no responses, and tradeoffs — and structuring them into a ranked holding queue — was critical to making the system feel usable and complete.
Users were more comfortable with AI when it showed what it checked, what it didn't know, and why it made a recommendation. Confidence increased when the system exposed its reasoning — even when it wasn't certain.
Expand the scheduling framework to additional workflows (appraisals, title, post-close). Introduce automated outreach (SMS / email) to reduce manual vendor contact. Enable real-time approval flows for fee and time exceptions. Continue evolving toward a fully agentic system that can act, not just assist.