Ava Liu

The Problem

A production resume agent grew to 99 context files, 38,924 provider input tokens, and a request lifecycle that crossed its 90-second boundary. Extending the timeout would only hide the architectural problem.

Approach

I measured context assembly, retrieval, provider usage, wall time, cost, and output quality separately, then redesigned the runtime around persistent retrieval and durable execution rather than request-time file parsing.

What I Built

• Persistent hybrid retrieval instead of request-time parsing
• Task-aware context budgets and prompt caching
• Quality guardrails and requirement-level validation
• Durable async execution with idempotency and failure classification
• Provider-authoritative token audits and observability
• Phase gates that blocked advancement when required audits failed

Why This Matters

The optimization became a production-runtime redesign: faster and cheaper only counted if factual recall and requirement coverage remained intact.

The Problem

Financial data is abundant, but decision context and uncertainty are fragmented. A single blended confidence score hid whether low trust came from thin data, weak analysis, or conflicting evidence.

Approach

I built a top-down Market → Sector → Industry → Stock decision structure, made uncertainty inspectable, then evolved the research layer into context-aware strategy ranking, parameter optimization, and leakage-controlled prediction validation.

What I Built

• Independent Data Quality, Analysis Reliability, and Evidence Alignment metrics
• Canonical Decision Engine across scanner and stock workflows
• Context-aware strategy ranking and stateful execution research
• Parameter search, objective-function design, and chronological out-of-sample validation
• Coded React/TypeScript DOM design system with semantic, ARIA-aware components
• Motion/GSAP interaction engineering, auth, Stripe subscriptions, admin, and Cloudflare deployment

Why This Matters

One product connects explainable AI, quantitative research, full-stack implementation, subscription infrastructure, and production design engineering without hiding uncertainty behind a single score.

The Problem

Support diagnosis is conditional: the same inbound email can require live account state, subscription and entitlement context, and proprietary product knowledge. A useful agent has to investigate without gaining uncontrolled authority.

Approach

I designed a bounded, model-directed tool loop with governed hybrid RAG, persistent execution traces, confidence-based escalation, and mandatory human approval before outbound communication.

What I Built

• OpenAI Responses API with multi-step function calling
• Read-only tool registry for account, subscription, entitlement, login, market, and product context
• Approved knowledge corpus with lexical + embedding retrieval
• Source, section, confidence, tool, and retrieval provenance
• Persistent conversation, draft, run, and audit state
• Resend inbound email, signed webhook verification, threading, Admin review/edit/send workflow

Why This Matters

The agent can investigate dynamically and produce grounded drafts while consequential customer communication remains explicitly human-controlled.

The Problem

A tailored resume is useless if generated claims cannot be traced to verified candidate evidence. Real usage exposed brittle lexical matching, stale inference, incomplete bilingual output, and completion states that could resolve before evidence verification finished.

Approach

I rebuilt the trust model around a canonical Evidence Library, semantic evaluation, source lineage, claim-level verification, human override, and independent review.

What I Built

• Canonical evidence model separating raw source, confirmed fact, interpretation, and resume-ready wording
• Semantic requirement/evidence evaluator with EN/ZH language invariance
• Source lineage, version history, and non-destructive evidence updates
• Claim-level trace persistence and independent factuality verification
• Independent ATS, recruiter, and hiring-manager review loop
• FastAPI + Next.js/React + TypeScript + SQLModel/SQLite + OpenAI API

Why This Matters

More than 90 logged end-to-end test rounds changed the architecture itself — including replacing brittle substring classification and repeated merge patches with semantic evaluation and a reconciled evidence model.

The Problem

An affiliate feature becomes a financial system as soon as attribution affects money. Identity, historical program terms, refunds, vesting, payouts, authorization, and auditability all have to stay correct over time.

Approach

I treated revenue state as backend-authoritative and historically reconstructable, with append-only financial ledgers, idempotent Stripe-derived processing, explicit reversals, and human-controlled payouts.

What I Built

• Affiliate applications, governed accounts, programs, campaigns, and referral links
• Click attribution, registration binding, and Stripe revenue-event ingestion
• Append-only commission and reversal ledgers
• Monthly-plan commission holds and annual-plan monthly vesting with per-tranche holds
• Manual payout recording, analytics, authorization isolation, and fraud/risk foundation
• Creator performance scoring without exposing referred-customer PII

Why This Matters

The system turns a growth idea into auditable commercial infrastructure: acquisition → conversion → revenue → commission → reversal → payout → risk → creator performance.

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/TypeScript design system, used Motion and GSAP for distinct responsibilities, and simplified a multi-clip interaction into one bidirectional media timeline.

What I Built

• Semantic DOM primitives and reusable product components
• ARIA-aware controls and component-level accessibility
• Motion for React microinteractions
• GSAP coordinated timelines
• Browser-native video state for interruptible interaction
• Reduced-motion behavior and production state handling

Why This Matters

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

The Problem

A 38% generate-to-preview rate was producing near-zero purchases. Interviews and usability testing showed the real blocker was not acquisition or pricing, but trust in the generated likeness.

Approach

I redesigned personalization, preview optionality, state continuity, payment identity, recovery, and fulfillment rather than treating model generation as the whole product.

What I Built

• AI preprocessing for distinguishing pet features, expression, and personality cues
• Structured personalization context feeding the generation workflow
• Three-preview strategy with parallel n8n generation
• Stripe payment tied to selected-asset identity
• Persistent fulfillment state and recovery paths
• Automated delivery with human recovery when needed

Why This Matters

The product improved by designing around model uncertainty: research identified the trust failure, and workflow architecture converted probabilistic outputs into a more dependable buying experience.

The Problem

A growing two-sided marketplace needed a consistent product and brand language across high-intent web journeys, financing interactions, marketing, and responsive experiences.

Approach

I built shared foundations, reusable responsive patterns, interaction standards, and conversion-oriented hierarchy so product, engineering, and marketing could work from the same system.

What I Built

• 0→1 brand and product design system
• Reusable responsive components and interaction patterns
• Marketplace and interactive financing UX
• CTA hierarchy and conversion-focused page structure
• WCAG/ADA-aware reusable patterns
• Cross-channel standards across web, social, email, and print

Why This Matters

The project shows that the coded design-system work in my AI products builds on earlier product-design practice, cross-functional implementation, and measurable business outcomes.

The Problem

Marketing needed to adapt source creative across platform sizes and aspect ratios without making designers a production bottleneck or losing brand consistency.

Approach

I translated spacing, hierarchy, composition, safe-area, and brand rules into a reusable AI-assisted workflow so non-designers could generate viable assets within defined constraints.

What I Built

• Weavy-based multi-format creative workflow
• Reusable spacing, layout, and aspect-ratio rules
• Brand-safe generative constraints
• Platform-specific adaptation logic
• Quality review checkpoints
• Team enablement for non-designers

Why This Matters

The workflow turns repetitive production knowledge into reusable operational logic while preserving human design judgment and brand standards.

The Problem

Generative tools could increase creative throughput, but inconsistent prompting and output quality could quickly fragment the brand.

Approach

I treated AI as part of a governed creative system: reusable templates, brand-safe prompting practices, asset libraries, quality benchmarks, and human art direction remained the control layer.

What I Built

• Brand guideline and presentation-template system
• Reusable asset libraries and production standards
• Midjourney, Firefly, and Runway integration into creative workflows
• Brand-safe prompting practices and defined use cases
• Visual quality benchmarks and review criteria
• Scalable templates for team self-service

Why This Matters

The project shows how generative AI can increase creative capacity without replacing the judgment, consistency, and governance expected from a mature brand system.

The Problem

Individual AI images or clips do not automatically become a coherent story. Character continuity, pacing, visual language, narration, and sound still need direction across the full sequence.

Approach

I used generative tools as part of an end-to-end creative production pipeline while retaining human control over concept, script, art direction, selection, sequencing, and final quality.

What I Built

• Original concepts, stories, scripts, and character development
• Midjourney and Runway visual exploration and generation
• Iterative character and scene continuity
• Animation, editing, timing, and transitions
• Narration, sound, and music integration
• Finished short-form storytelling releases

Why This Matters

The series preserves the creative-direction side of my practice: AI accelerated asset creation, but story, taste, continuity, and final editorial decisions remained human-led.