Multi‑disciplinary working sessions

• Key moments of experience and ideal scenarios
• Front‑stage: Patient‑facing interactions, underlying assumptions, and key questions
• Back‑stage: Internal processes, identified assumptions, gaps, and inquiries
• Ideal solution, technical (currently digitized, partial, or a gap) assumptions, and questions

Pre‑arrival

Imagine a pre‑arrival experience that feels guided and empowering. The system orchestrates everything ahead, seamlessly and transparently prompting the patients with the right steps, so they and the care team feel informed and prepared for the upcoming visit.

Arrival and rooming

Imagine that the patient would always know exactly where to go and how to get there, and have their arrival acknowledged immediately, so they feel seen, expected, and known.

In‑between visits, day‑long coordination

Imagine a patient day unfolds with clarity—where they're gently guided through what's next, kept informed as plans evolve, and able to move through appointments with confidence, not confusion.

Transformative rooming: sample of ideal experiences

• Patients know where to go and how to get there and their arrival is acknowledged immediately.
• The system seamlessly checks patients in for all their appointments and their identity is confirmed without asking them to repeat their name or date of birth.
• Patients understand what's happening next, and why and know how to get updates if things change.
• Patients know when their room is ready and how to get there and when they arrive, the care team member they need to see is already there for them.

Moment 3: Welcomed, right on time

FROM

• Patients are often confused when navigating the campus or moving between departments.
• Patients arrive 10–15 minutes early and join the queue to check in at the front desk, kiosk, or through verbal confirmation.
• The system cannot take advantage of patients who arrive early, since the current process is strictly sequential.

TO

• Clear guidance confirms they're on the right path, and arrival options appear automatically, tailored to accessibility needs and visit context. The system recognizes the arrival and signals to both patient and care team that the visit has begun. There's no need to announce arrival; it's already known.
• The patient feels received, oriented, and ready, like a guest whose host has been waiting.

Moment 7: Guided through the day

FROM

• Resident patients may wait either in an exam room or in the waiting area for additional procedures. For some, the initial room is not their final destination for the day, as their visit involves multiple steps or locations.
• For patients with multiple appointments in a single day, the same steps and challenges repeat at each visit: they must check in separately for each department, may need to fill out some questionnaires manually if not available on the portal, and are repeatedly asked to provide or verify their identity, insurance, and demographic information.

TO

• Time is respected by design — clear readiness signals and coordinated transitions maintain momentum across the day.
• Keeping patients on schedule and providers aligned ensures visits move predictably toward on‑time departures.

Experience meter

WHAT

Long‑running agent with broad, read‑only access to internal systems to identify bottlenecks and notify care team for validation and triage.

WHY

Variety and complexity of systems at play means that automation would be complex to implement and maintain. GenAI Agent with tooling could dynamically identify issues across many tools, domains, and processes.

Recommendations

Helped shape the feasibility plan to ensure UX actively informed technical exploration, using experience research, co‑design, and mapping to guide data readiness, integration feasibility, and Responsible AI considerations. This approach reduced risk early by grounding technical decisions in real ambulatory workflows and clarifying what could be safely tested in a Minimum Viable Experiment versus what required further validation.

Getting to know care managers

Their work involves conducting lengthy health risk assessments that can take upwards of two hours per session, navigating multiple systems to piece together each member's medical history, and ensuring every piece of documentation meets strict National Committee for Quality Assurance (NCQA) compliance standards.

Despite the heavy administrative burden — from detailed note-taking to validating assessment notes across roughly 1,800 assessments per month — their core mission remains deeply human: to build trusting relationships with members, to be fully present and attentive during interactions, and to streamline the administrative weight so they can focus on what matters most — truly listening, understanding, and advocating for the people in their care.

Health risk assessment

Care managers are required to complete a Health Risk Assessment (HRA) for each member — a standardized form consisting of approximately 150–200 questions spanning multiple input types such as checkboxes, radio buttons, and free‑text responses.

Accurate completion of the HRA is essential, as it informs care planning decisions and supports regulatory documentation — making it a necessary but administratively intensive component of the care manager's workflow.

Main pain points

• Spending a significant amount of time on detailed note taking and completing lengthy assessments, which can be overwhelming and time-consuming.
• Multiple systems to look up member information including paperwork from doctors and nursing homes.
• Engaging members during the session and building trust with limited time with the members.
• Ensuring all documentation meets NCQA requirements can be stressful and time consuming.

Role of AI

AI could reduce administrative burden, but in this context, it could just as easily:

• Erode clinical judgment
• Create false confidence in incomplete data
• Shift responsibility away from care managers
• Fail regulatory and ethical review

My contributions

As the Principal Designer, I shaped the experience, strategy, and Responsible AI approach across discovery, experimentation, and roadmap definition, working closely with care managers, data science, engineering, and governance partners.

• Framed the problem and narrative around care manager presence, judgment, and accountability
• Translated research, regulation, and AI uncertainty into design decisions
• Co-defined ground truth, evaluation metrics, and human‑in‑the‑loop feedback loops for safe AI
• Co-designed interaction patterns that encourage oversight, not over‑reliance
• Co-defined a strategic path for AI adoption

The design approach

1. Human-centered design from the start
We co‑designed directly with care managers, observing real assessments, mapping mental models, and making AI tangible early through low‑fidelity concepts and provocations.

2. Systems-level thinking
I treated the experience as a system, not a screen, connecting front‑end interactions to transcription quality, data readiness, NCQA requirements, and AIRB review.

3. Responsible AI as design material
Responsible AI wasn't a checklist. It shaped what we built, what we delayed, and what we refused to automate.

My contributions

• Framed the right problem for a high‑stakes clinical AI system and champion a human-centered design approach throughout the project
• Established a cross-disciplinary operating model that brought clinicians, client stakeholders, and internal teams together to make aligned, system-level decisions
• Co-designed the end-to-end physician experience and product interactions
• Generated human-centered success metrics beyond accuracy
• Ensured safety, trust, and adoption in real clinical workflows

Grounding the work in clinical reality

Our approach to understanding the problem involved conducting exploratory research and engaging in field visits with physicians, allowing us to observe actual admission workflows, interruptions, and the pressures of time firsthand.

We carefully mapped the current methods of assessing VTE risk, examining the various tools, documentation, and mental shortcuts used.

From the very beginning, we developed concrete AI concepts to test our assumptions and provide clinicians with tangible solutions to respond to.

Insights

• Physicians must search across vast amounts of unstructured EHR data, notes, discharge summaries, and historical records—where relevant VTE risk information is hard to find, inconsistently documented, and easy to misinterpret.
• Clinical accuracy is not only a function of correct data, but of whether clinicians can find and trust the right information within severe time constraints.
• Even a single 'yes' can be decisive in physicians' decision to administer or withhold VTE prophylaxis.
• Physicians already have a concept of their patients, so if AI surprises them, they'll investigate further.

Workstreams overview

The workstreams worked in parallel and built on each others' learnings.

• Human & AI Experience Design drove requirements gathering which led to experimentation validating those requirements and generating hypothesis based on them. User design preferences were directly implemented in application.
• Validated experiments were replicated in the orchestrator and surfaced in the application. Findings of experimentation inspired further discussion with the clinicians.
• Application usage and feedback inspired further discussions with users, and iteration over requirements as well as testing further improvements in experimentation.

Designing for workflow, not screens

• Co-designing and iterative experience prototyping of critical decision moments: AI suggestions for 20 risk indicators, citations, reasoning, failure modes, feedback loops, and onboarding experiences
• Co-defining experiments and project success metrics
• Mapping end‑to‑end clinical workflows, not just UI states or happy paths to see how the AI fits into the admission flow, interruptions, and time‑compressed decisions.
• Making risk explicit and measurable: Failure modes, Over‑reliance risks, & feedback and correction loops

Human and AI experience experiments

• How might we (HMW) create the most appropriate info architecture that meets the physician needs and accommodates the different data categories and related interaction patterns?
• HMW define the solution success systematically, starting with what is desirable for the physicians?
• HMW ensure a holistic approach where the solution is meeting the needs of different use cases and not only the happy path?
• HMW co-create safeguards that prevent harm and support user control?
• HMW empower users to review, contextualize, and edit the AI-generated content?
• HMW intentionally design for the moments of failure?
• HMW gather feedback from the physicians in a meaningful and practical way?

HMW gather feedback from the physicians in a meaningful and practical way?

Goals: user sentiments, where we made a mistake to improve the results, enrich the ground truth

Requirements: For the pilot it might be ok to add the feedback for all the risk indicators and citations but for a larger scale launch, feedback should be collected on a broad level rather than for each individual risk indicator, which simplifies the process:

"Give people a chance to add their open-ended feedback if they want to capture more detailed and valuable feedback, people be able to explain their ratings and provide specific suggestions for improvement." —Physician

Championing a responsible approach

Fostering human-centered AI conversations: Relying on the HAX toolkit, Overreliance framework, and other RAI best practices to create materials to inspire cross-disciplinary collaboration

Setting the right foundation: Co-creating the impact assessment and other documents in response to ORA recommendations, and translating those to client-facing and design materials, including providing control for care managers

Bringing contextual knowledge through research: Uncovering care managers needs, concerns, and desires, and what they want to know about AI to inform use cases, disclaimer content, training materials, datasets, and evaluation criteria

Impact

Clinical workflow impact: Reduced manual chart review and improved access to risk indicators in one place, as validated through clinician feedback

Business impact:

• First-of-its-kind CDS with GenAI at scale
• 110,000 admissions annually impacted
• Enhanced care quality with strategic and financial impact
• Enabled expansion beyond VTE into additional clinical domains using the same core UX and engineering foundation.

The JHM VTE project was selected for Engineering the Amazing, recognizing it as a standout example of co‑engineering, responsible AI, and real‑world clinical impact

Read the scientific paper (medRxiv)

My contributions

1. Grounded the team in the real needs & desires of Maya artists rather than creating a product from a technology capability perspective only.
2. Led the team to a collective strategic direction that provided clarity, alignment, and commitment.
3. Brought stakeholders together through a participatory, transparent, and responsible process.

Artist profiles

• Novice artists: Takes a methodical and measured approach to learning Maya skills, emphasizes "pacing yourself" to avoid getting overwhelmed.
• Professionals: Believes that artists apply principles and theory to their work to create beauty and believability, and that AI is unable to do so.
• Educators: That Maya is incredibly extendable in ways other platforms are not. One just needs to learn the right skills.

Decision-making criteria

• Desirability (Valuable to Maya artists)
• Project time to value (Required time to have a working capability for the artists)
• Technical complexity (Difficulty to implement, level of engineering effort)
• Enhancement potential (Adaptable to future enhancements, reusable)

Creative facilitation

Fostering active participation, promoting cross-perspective learning, soliciting feedback, and effectively conveying messaging.

• Idea generation sessions with the internal and client teams to bring to life potential benefits and failures of the technology
• Co-design workshops with Maya artists to gain feedback about the generated concepts in the collaborative sessions with the internal and client teams as building blocks to envision an ideal future
• Strategy workshops to decide on the decision-making criteria and prioritize opportunity areas

Solution

FROM

Make our technology work for Autodesks' Maya

TO

Imagine how a product experience could be enhanced when intelligent technologies anticipate what Maya artists need

Background

Pancreatic cancer is one of the most aggressive and time-pressured diseases in oncology. Treatment paths diverge sharply and early - surgery first or chemotherapy first, which regimen, how aggressively - and the consequences of the choice are severe. To support these decisions, the data-science team built an unsupervised machine-learning model that groups historical patients into clusters of clinically similar cases and lets a physician see how patients "like mine" were treated and how they fared.

The modeling was strong; the experience was not. The first build exposed the machine learning directly to the clinician: a principal-component scatterplot with statistically-labeled axes, clustering parameters and feature tags as raw controls, and dense small-multiples of box plots and pie charts. It answered the data scientist's questions, not the doctor's.

My Contribution

1. Framing the Problem & Clinical Context

Reframed the product from exposing model outputs to answering clinician questions, grounding the solution in real workflows across initial and longitudinal patient use.

2. Participatory Co-Design

Led an iterative co-design process with clinicians and data scientists, using rapid prototyping and exploratory concepts to converge on a clinically meaningful MVP.

3. End-to-End Experience Design

Designed the connected system across first-visit and returning-patient flows, establishing the information architecture and comparison patterns that maintain continuous patient context.

4. Explainability & Responsible AI

Integrated explainability and Responsible AI by defining clear language, interaction patterns, and guardrails, surfacing uncertainty, data provenance, and "context, not prediction" cues to support trust without overreliance.

Getting to know the oncologists

Through secondary research and working sessions with our clinical partner:

• Time is brutal: appointments run 10-15 minutes, and clinicians spend roughly 30-40 minutes preparing for each new patient, much of it reconstructing comparable cases from memory.
• Mixed expertise: about 30% of patients are seen by nurse practitioners, so the tool cannot assume deep subspecialty fluency or tolerate unexplained jargon.
• Two real moments of use: preparing before a visit, and sharing context with the patient during it.
• Trust is the gating factor: clinicians described the model as a "black box". If they cannot see why a patient maps to a group, they will not act on it. Interpretability, not accuracy, was the adoption bottleneck.

Patient Journey

Pancreatic-cancer patients do not follow one path, so the tool could not assume one. Working from the clinical team's input, I mapped the journey the design had to support:

• First visit and staging: the patient is assessed and routed along a metastatic (META) or non-metastatic (NON-META) path.
• Non-metastatic patients split further: resectable, borderline resectable, or locally advanced, driving an upfront-surgery versus chemotherapy-first decision.
• Treatment proceeds along standard regimens: for example, FOLFIRINOX, or gemcitabine with nab-paclitaxel, with response and tolerability tracked over time.
• Recurrence loops the patient back: into re-assessment and a new decision.

This journey became the backbone of the experience: a first-visit "new patient" flow for orientation, and a "returning patient" flow centered on trajectory, how this patient is doing relative to similar patients over time.