Ava Liu

The Problem

Market data is abundant, but decision context is fragmented. Users need to understand the environment, where strength is concentrated, whether a stock is actionable, and what evidence supports the conclusion.

Approach

I designed a top-down decision path and built uncertainty, provenance, and role-aware SaaS behavior into the architecture rather than treating them as UI decoration.

What I Built

• Three independent trust metrics
• Canonical Decision Engine across Scanner and Stock DNA
• Recorder / Replay / Learning architecture
• Auth, Stripe, entitlements, admin, beta, analytics
• Coded DOM design system and motion engineering
• Cloudflare Tunnel — secure self-hosted ingress, no public IP

Why This Matters

A single project spanning product strategy, XAI, full-stack implementation, SaaS operations, data systems, interaction design, and design engineering.

The Problem

Support diagnosis is conditional. The same email can require live customer/account evidence and proprietary product knowledge.

Approach

I evolved the system into a bounded model-directed loop and added governed hybrid RAG, while preserving explicit human authority over outbound communication.

What I Built

• Multi-step function calling
• Read-only tool registry
• Approved knowledge corpus + hybrid retrieval
• Persistent tool and retrieval traces
• Thread-aware conversation state
• Admin-gated Send
• Dedicated inbound email infrastructure — isolated subdomain, DNS/MX config via Cloudflare

Why This Matters

The system can investigate autonomously without being allowed to take consequential customer actions.

The Problem

Motion had to add delight to a dense financial product without becoming a one-off landing-page trick or breaking state integrity under fast pointer input.

Approach

I extended the coded React design system, used Motion and GSAP for distinct responsibilities, and simplified the mascot interaction to one bidirectional media timeline.

What I Built

• Semantic DOM primitives
• ARIA-aware controls
• Motion for React
• GSAP timelines
• browser-native video state
• reduced-motion behavior

Why This Matters

Visual craft translated into semantic components, accessible behavior, animation architecture, and real browser-state control.

The Problem

A tailored resume is useless if generated claims cannot be traced to candidate-confirmed evidence. Real usage exposed brittle lexical matching, stale inference, and verification gaps.

Approach

I rebuilt the trust model around source provenance, semantic evidence evaluation, canonical evidence, version lineage, and independent review.

What I Built

• Semantic requirement/evidence evaluator
• Bilingual evidence workflow
• Source Library + version history
• Claim-level verification
• Independent ATS/recruiter/hiring-manager review

Why This Matters

Real end-to-end testing changed the architecture rather than only changing prompts.

The Problem

Analytics suggested a funnel problem; interviews showed the actual blocker was likeness and trust in the generated result.

Approach

I redesigned inputs, optionality, state continuity, payment identity, and recovery instead of treating model generation as the whole product.

What I Built

• Input normalization
• Pet-specific personalization
• Three-preview optionality
• Stripe selected-asset identity
• Human-in-the-loop recovery

Why This Matters

Reliable AI commerce comes from system design around uncertainty, not pretending the model is deterministic.

The Problem

Marketing needed to produce more campaign content without making designers a production bottleneck.

Approach

I translated brand and layout standards into reusable AI-enabled workflows and enabled non-designers to use them safely.

What I Built

• Structured prompting
• Visual guardrails
• Multi-format resizing
• Reusable templates
• Team mentoring

Why This Matters

The technical portfolio is built on a long foundation of visual judgment, storytelling, and creative leadership.