RASHMI
PADMANABHA

Growing mobile traffic, rising accessibility litigation risk, and new AI capabilities created the case for a full redesign. Six things were broken, and each one had a real cost attached:

• Mobile-first failure — 48% of traffic was mobile but the UI was desktop-designed. Navigation collapsed, text overflowed, touch targets undersized.
• Accessibility risk — 47 WCAG failures at baseline (2.1 AA audit): colour contrast, broken focus states, screen reader semantics. An ADA risk and a real user blocker.
• Findability crisis — no unified search. Users averaged 4.2 clicks to find a relevant article; the symptom checker was effectively invisible.
• Content overload — long-form articles with no chunking. 62% abandoned before reaching actionable guidance, above the 40% norm.
• EHR integration gaps — clinician-sent education links delivered generic content; condition-specific linking broken across 3 EHR partners.
• No AI advantage — a vast content library but zero personalisation, zero reading-level adaptation, zero AI-assisted discovery.

Sequencing: heuristic evaluation + analytics first (to brief stakeholders fast) → stakeholder interviews → user sessions → accessibility audit → competitive benchmark.

Card sort finding: an open card sort (Maze, 48 participants) revealed 3 distinct mental models — condition-first (62%), symptom-first (28%), topic-first (10%). No participant grouped content the way the original IA was structured; "Health Library" was recognised by only 34%. "Topics" replaced four ambiguous, overlapping nodes ("Health Library", "Wellness", "Diseases", "Conditions") with one condition-first entry and three clear sub-categories.

How Might We — the five questions that framed ideation:

• Make health content discoverable in under 2 taps on mobile?
• Use AI to personalise content to a user's condition and literacy level?
• Give clinicians a frictionless way to assign relevant education?
• Surface the symptom checker as the primary entry point?
• Make long articles feel manageable without losing depth?

Prioritized on effort vs. impact — the high-impact, low-effort quadrant shaped the first release:

• AI-Assisted Content — GPT-powered reading-level adaptation, personalised article summaries, AI-driven recommendations based on condition, history, and literacy.
• Unified Smart Search — single search with condition/type/literacy filters, autosuggest, contextual ranking. From 4.2 clicks to under 2.
• Mobile-First Redesign — bottom nav, swipeable topic cards, 44px+ touch targets, progressive disclosure. Built mobile-out, not adapted down.
• Accessibility by Design — WCAG 2.2 AA token system embedded in the design system. Zero critical failures on final audit, ADA + EAA compliant.
• EHR Content Linking — contextual deep-linking from clinical workflow. Condition-matched article assignment in one click from within the EHR.

Six key flows wireframed lo-fi first — symptom-to-content, search-to-read, EHR-link-to-article, homepage-to-article, reading-level adaptation, and save/return/progress — tested in 2 rounds (n=12) before moving to hi-fi.

• Lo-Fi (FigJam + paper) — rapid sketches of 6 flows, IA decisions validated, 2 rounds/12 users. Caught 4 critical navigation errors early.
• Mid-Fi (Figma, grayscale) — clickable prototype focused on interaction patterns: search filters, accordion behaviour, reading-level toggle. Task completion: 73%, up from a 51% baseline.
• Hi-Fi (Figma, branded) — full visual fidelity with the design system applied, 20-user validation. Task completion: 87%, 4.8/5 ease score.

What testing changed:

• Bottom nav reduced task time by 42% on mobile vs. a hamburger menu.
• Unified search with autosuggest and condition filters lifted search-to-article success from 44% to 82% — patients found the right article far more often on the first try.
• AI summary card increased full-article reads to 71% of the time.
• Reading-level toggle used by 68% of low-literacy participants — unprompted.
• First search filter iteration had too many options — simplified from 8 to 3.
• EHR link flow needed 2 additional confirmation steps after clinician testing.
• Symptom checker awareness rose to 83%, up from 19% pre-redesign.

• Colour contrast — all text ≥ 4.5:1, large text ≥ 3:1, no colour-only indicators, tested across light and dark surfaces.
• Keyboard navigation — full site keyboard-operable, logical focus order, custom 3px teal focus rings, skip-to-content on every page.
• Screen reader support — semantic HTML5 + ARIA throughout, live regions for dynamic AI content updates. Tested with NVDA, JAWS, VoiceOver.
• Reflow & zoom — content reflows at 320px (400% zoom), no horizontal scroll, touch targets ≥ 44×44px on all interactive elements.
• Readable content — plain-language guidelines, Flesch-Kincaid score ≤ Grade 8 default, reading-level selector, tooltip definitions for medical terms.
• Compliance result — 47 failures at baseline (2.1 AA audit) → 0 critical failures at final audit. WCAG 2.2 AA achieved, ADA compliant, EAA 2025 ready, verified by third-party audit.

• Wearing both hats sharpens prioritisation. Being PM and PD forced ruthless decisions about what to design — every choice had a business case behind it. Less decoration, more function.
• AI doesn't replace design — it expands the design space. Integrating AI content features meant designing for uncertainty: summaries that might be wrong, recommendations that needed guardrails. AI raised UX complexity, not lowered it.
• Accessibility is a PM problem as much as a UX one. Getting a11y into sprint planning, audit timelines, and EHR partner sign-off was product management work — design alone doesn't ship accessible products.
• Test with the hardest users first. Running round-1 sessions with low-literacy users meant design assumptions were challenged while it was still cheap to change them — it saved weeks of rework.

The process was as risky as the label itself: personal naming conventions stood in for version control, AI flags carried no rule or confidence so reviewers closed every one, approvals moved by email reply with no audit trail, and print operators received files with no confirmation of approval status.

Data gathered: 80+ people across three role groups, 400+ data points across two usability rounds, and a full mapping of exactly where the artwork "lost its identity" between handoffs.

Insight → Decision: Users had already workarounded the broken system into something simpler — the real workflow was invisible inside three disconnected tools.

Scored against the personas' pain points — Impact on trust & compliance risk vs. Effort to ship.

The product runs on Roboto throughout — chosen for its neutral, highly legible letterforms and tabular figures, which matter when reviewers are scanning version numbers and rule codes under time pressure.

• AI trust is earned through transparency, not capability.
• Design the shared data model before the first screen.
• The last person in the chain defines whether the system worked.

• Visitor / Client — needs fast entry, clear guidance, and a premium hospitality experience.
• Host / Employee — needs visit planning, tracking, room readiness, and coordination visibility.
• Receptionist — needs reduced manual workload and better prioritisation during peak times.

Data gathered: Logged reception overload during visitor peaks, host follow-through on food/room/access arrangements, and where clients lost confidence in the flow after check-in.

Insight → Decision: Check-in was not a single kiosk interaction — it was a chain of dependent moments across people, access, space, and hospitality. This reframed the design problem from "screen design" to system design.

• System design was more valuable than isolated screen design — the experience depended on multiple teams acting at the right time.
• Automation reduced cognitive load for visitors, hosts, and receptionists alike.
• End-to-end thinking made the experience feel premium, because support continued after the check-in moment.

Insight → Decision: After empathising with users, their needs, pain points, and frustrations doing daily activities became clear — different people had different needs, which a single portal redesign wouldn't solve. That pointed toward a conversational assistant available anywhere, not another screen to navigate.

Scored for immediate release against the full backlog of requested features.

• Meeting employees inside a tool they already have open all day removed more friction than any portal redesign could.
• Designing directly in the bot framework's own tooling surfaced platform constraints early, before they became rework.
• Close collaboration with developers was essential to shape a good experience around the real limitations of a chatbot.

Data gathered: Observed employees performing common tasks to surface pain points and confusion, paired with survey data for measurable, quantitative validation of what the sessions showed.

Key findings:

• Navigation issues — employees struggled to locate apps and information in a complex structure.
• No responsive design — the portal wasn't optimized for mobile, blocking access on the go.
• Inconsistent layout — poor visual hierarchy made the experience confusing and overwhelming.
• Limited access to information — critical content like policy benefits was buried deep in the IA.

Employees were invited to perform specific tasks on the redesigned portal while I observed their interactions, collected feedback, and measured task completion time and success rates. Results were positive — the simplified navigation, mobile optimization, consistent layout, and search functionality streamlined processes and improved employee satisfaction.

• Session analysis paired with survey data gave both the "why" and the measurable "how many" — neither alone was enough to prioritize confidently.
• Grouping by task, not by system or department, is what actually reduces navigation confusion.
• Mobile access isn't a nice-to-have for a self-service portal — it's where a large share of quick tasks happen.

• Director of Assets — portfolio-wide KPIs, Decent Homes Standard score, cost analysis.
• Void Manager — void pipeline by stage, re-let time vs target, rent loss exposure.
• Compliance Officer — certificate status, overdue escalation flags, regulatory risk.
• Repairs Manager — work order queue, SLA status, contractor assignment.

Data gathered: Measured average void re-let time against target SLA, tracked how often compliance deadlines were missed due to lack of visibility, and mapped duplicate data entry happening independently across teams.

Insight → Decision: All four roles needed the same underlying data but conflicting views of it — so the decision was one shared data model with role-based dashboard routing, instead of building four separate apps.