Good bones, generic paint.

A layered ::after-overlay interaction layer driven by design tokens, wired cleanly into the button and badge primitives. Genuinely sophisticated — most shadcn clones never build this. It is the app's real design signature.

Primary is a professional blue (hsl(210 85% 45%)) on every logo lockup and CTA — the deliberate choice that keeps the app off the AI-default palette. This is the single biggest thing protecting MAPLE's first impression.

When comps are missing, the app refuses to render MAO / score and shows an amber explainer instead of a misleading number. This is exactly the judgment a valuation tool must have — never fabricate precision. Do not "simplify" it away.

A real content-shaped loading skeleton (not a spinner), and the one row primitive that gets numbers right: label left, figure right-aligned in font-mono tabular-nums. This row is the template every other numeric row should copy.

Risk / tax metrics expose their exact formula and substituted numbers on hover — load-bearing information, not decoration; the strongest crafted feature in the app. The AI chat sanitizes with DOMPurify + an explicit tag allowlist, escaping before markdown conversion. Correct and secure.

Four explicit states — loading / success / error / expired — each with its own icon, color, and recovery action. Most apps ship one generic screen; this is premium-grade. The auth flows overall (validation, resend cooldown, privacy-safe reset copy) are already at a good bar.

The mesh-background blobs are literally violet (hue 260° / 280°) — the exact purple-gradient-blob signature — and they clash with the blue brand. Used on landing, login, signup, onboarding, and more.

Fix → drop mesh-background for a flat bg-background (Linear/Stripe auth pages are flat), or retune all three blob hues into the brand's blue (205–225°) at ≤0.08 opacity. Two pages already use a flat background — make that the standard.

The base Card hard-codes the glass class, so every card gets a 16px backdrop-blur. That is the glassmorphism tell applied globally — and a real performance cost with dozens of blur layers on a data-dense page.

Fix → remove glass from the base Card; make it an opt-in variant="glass" used only on overlays / sheets where a blur behind actually reads.

The assistant is a wall of from-purple-600 to-indigo-600 gradients, purple avatars, and purple hover text — while the entire rest of the app is blue. It reads as a bolted-on ChatGPT clone with its own identity rather than part of MAPLE.

Fix → recolor every gradient to bg-primary / text-primary / bg-primary/10. Same code, and the assistant suddenly reads as MAPLE's own strategist.

The fetched-comps table renders price, sqft, beds/baths, DOM, and distance as default left-aligned, non-tabular text — visibly worse than the better comps table directly below it, and this is the first table users see after generating comps.

Fix → add text-right font-mono tabular-nums to numeric cells, text-right to the headers, and render price through the shared formatCurrency.

🚗 and 📍 emoji stand in for icons in the comps panel; the Sparkles glyph is sprinkled across six screens as a generic AI signifier — both classic slop tells in a financial UI.

Fix → swap the emoji for Lucide <Car> / <MapPin> (already imported); keep Sparkles on at most one primary AI entry point and use domain icons elsewhere.

Five copies of formatCurrency, two byte-identical score-color helpers, and three hand-rolled step indicators with different sizes. On a tool whose entire value is number consistency, this is a real divergence risk.

Fix → move formatters into the existing lib/numberFormat; extract one shared <Stepper>.

"Institutional grade", "the same tools institutional shops use", "10X your underwriting speed", and a stale "launching February 2026" — hype claims plus spec numbers relabeled as "social proof."

Fix → soften to defensible language ("institutional-style comps scoring", "analyze deals in minutes"), drive the date from config, and relabel the spec tiles honestly — never dress product specs as adoption.

An infinite @keyframes ai-glow in literal rgba(147, 51, 234) — the exact violet the research calls out — pulsing forever on the floating AI button and the voice button. The textbook "AI thing must glow purple forever" tell.

Fix → delete the keyframes, the utility class, and both usages. A static shadow-md is enough.

The Comps tab renders the same comp set twice — the poorly-formatted inline table (flagged above) and, ~40 lines below, the richer panel with $/sqft, score, condition, and selection. Users scroll past two tables of the same data. Highest-value single cut: it removes redundancy and the worst-formatted table at once.

Fix → delete the inline table (and its mobile twin); keep the richer panel.

Two unused onboarding step components, an unused duplicate gauge, undefined animation-delay-150/300 classes (which is why the "staggered" pulse dots actually blink in unison), and a raw JSON tab shown to end users as a first-class view.

Fix → delete the unused files and no-op classes; move the JSON dump behind an admin flag.

New users hit a dashed-border "Video walkthrough / Coming soon" placeholder on the very first screen — a dead deliverable stub in the highest-visibility spot in the app.

Fix → remove the placeholder until a real video exists; keep the component for when a URL is supplied.

Zero prefers-reduced-motion rules anywhere, yet the app runs infinite glow, bouncing typing dots, and pulsing loops. A WCAG 2.2 (2.3.3) gap and a polish gap.

Fix → add one global @media (prefers-reduced-motion: reduce) block that neutralizes the looping animations.

The valuation confidence is computed but never shown — the "Choose ARV" card displays only a source badge. For a tool people cite to clients, decision-critical numbers should carry a Low / Med / High confidence badge and an "as of" timestamp.

Fix → badge confidence next to the anchor number; add a comp-set freshness date.

Neither comps table lets you sort by $/sqft, distance, DOM, or score — the four fields an underwriter actually sorts on — and scrolled table headers disappear.

Fix → clickable sort headers with an indicator; sticky top-0 on the header row.

Icon-only tooltip triggers are bare SVGs with no tabindex — mouse-only. Many custom buttons lack a visible focus-visible ring. And if an AI message fails to send, the user gets silence with no retry.

Fix → wrap icon triggers in a focusable <button aria-label>; add focus-visible:ring-2 ring-primary globally; add an inline error/retry row to the chat.

KPI cards show a static number and a static icon; the monthly-trend data is already fetched but unused — a sparkline or "+N this month" delta is what separates an analytics dashboard from a counter. And only one component uses tabular-nums, so big figures jitter when they update.

Fix → add an inline sparkline / delta to the KPI cards; add tabular-nums to every mono figure class.