Ava Liu

The Problem

Most AI products present a conclusion and ask the user to accept it. In a domain where a wrong or misread signal has real consequences, that's not good enough. The core question: how does a user know when to trust an insight — not "is the AI usually right," but "can this user, looking at this output, tell whether to act on it right now."

Approach

Owned end-to-end product strategy, UX, information architecture, and AI system design for an explainable AI financial data analytics platform. Split a single blended confidence score into three independent metrics — Data Quality, Analysis Reliability, and Evidence Alignment — so a user sees not just how confident to be, but confident about what.

What I Built

• A multi-page information hierarchy turning high-density data into a structured, interpretable path
• Interactive, zoomable visualizations — time-series charts and multi-metric breakdowns across 180+ data points
• A personalized feedback layer plus a companion admin console for platform-level oversight
• An application-based beta access flow, balancing controlled rollout with a low-friction path in

Built With

FastAPI, Next.js, Python, Canvas-based rendering — directed via Claude Code and Codex.

Why This Matters for Forward-Deployed Work

This is the forward-deployed pattern end to end: sit with the real question a user needs answered, architect the system that answers it — reasoning pipeline, trust scoring, data visualization — and ship the full stack alone, with reliability built in for the moments the data can't support a confident answer.

The Problem

A resume-tailoring tool is only useful if what it generates is true. Early testing surfaced a harder question than "does the AI produce a good draft": can the system prove, claim by claim, that every fact in a generated resume traces back to something the candidate actually confirmed — and catch itself before it tells the user a document is done when it isn't.

Approach

Built and tested the product end to end against a real job posting instead of relying on unit tests alone — analyze → requirement map → gap interview → answer → re-analyze → tailored resume — logging each failure as it surfaced across 90+ rounds of live iteration spanning UX, data governance, and AI-reasoning layers.

What I Built

• The product's reasoning runs entirely on LLM calls (OpenAI API) — routing requests across model tiers by task complexity, and debugging real model-behavior failures (JSON-mode response errors, inconsistent output structure) before they reached the user
• A semantic evidence evaluator that replaced brittle substring matching — judges requirement intent, context, and transferability, with EN/ZH language-invariance so equivalent facts in either language score the same
• A four-layer trust architecture (canonical source, canonical evidence, generated-resume version, claim-level trace) so a completion state can't lie about what's actually been verified
• A Source Library with logical version lineage — replacing a resume creates a new version of the same source instead of false independent corroboration, while preserving full audit history
• A global-reconciliation redesign after repeated local fixes kept failing the same class of bug: replaced per-item AI confirmation with one verified-by-default Evidence Library, so users review exceptions instead of arbitrating every AI decision

Built With

FastAPI, Next.js/TypeScript, SQLModel/SQLite, OpenAI API — directed via Claude Code and Codex.

Why This Matters for Forward-Deployed Work

This project is the forward-deployed loop compressed into one build: run the system against a real case, let production usage — not a spec doc — surface what's actually broken, understand why the underlying LLM behaves the way it does, and fix root causes instead of symptoms. The independent ATS/recruiter/hiring-manager review layer plays the role a real customer would: an outside check that keeps the system honest about whether it actually solved the problem.

The Problem

A self-built, full-stack AI art product had a 38% generate-to-preview conversion rate paired with near-zero purchases. The easy read: an acquisition or pricing problem. Instead of assuming, I went to the users.

Diagnosis

User interviews and usability testing with non-converting users — not analytics alone — surfaced the real blocker: likeness. The AI-generated pets didn't look enough like the actual pet. This reframed the problem from "funnel" to "trust in the output."

Design Response

• An AI pre-processing step that converts each pet's distinguishing features into a personalized generation prompt
• Redesigned the preview flow from one image to three, engineering a parallel workflow so generation time didn't increase
• Iterated the loading experience from a generic bar to a step-by-step status display, keeping users patient through longer waits
• Built delivery logic to track and hand off the exact image a customer selected through payment

Why This Matters for Forward-Deployed Work

The fix came from going to users directly instead of trusting analytics, then engineering the production pipeline — parallel generation, a personalization pre-processing step — that made the fix real rather than just documented. Shipping a generative-model product meant designing for the model's actual behavior — a single generation is inconsistent by nature — rather than the behavior you'd wish it had. Diagnose with the user, architect the system, ship it: the same loop a forward-deployed engineer runs inside a customer's environment.

The Problem

Marketing teams needed to produce more campaign content across channels and formats, but relying on designers for every asset created production bottlenecks. Generative AI could increase speed, but without a structured system it also introduced inconsistent visuals, off-brand outputs, and unpredictable quality.

Approach

Designed an AI-enabled creative workflow that turned simple marketing briefs and variables into structured prompts, then supported image and video generation, brand-safe quality control, and automated adaptation across multiple formats. Built the workflow around reusable design rules and guardrails so non-designers could create campaign-ready assets with greater independence while maintaining brand consistency.

What I Built

• An AI creative workflow that translated simple marketing briefs and variables into structured generation prompts
• Repeatable image, video, and creative-generation steps with brand, layout, spacing, and quality rules built into the process
• An AI-powered multi-format resizing workflow (Weavy) that adapted creative for different marketing channels and formats
• Reusable templates and guardrails that enabled non-designers to generate campaign-ready assets independently
• Brand-safe prompting practices and quality benchmarks governing AI-generated visual assets
• Mentored the marketing team on prompting, workflow usage, quality control, and building practical AI-enabled creative processes

Outcome

Reduced dependency on designers for routine production by giving the marketing team a structured, brand-safe way to create and adapt campaign assets independently. Shifted AI from an individual creative tool into a repeatable team workflow. It's the same forward-deployed motion applied to enablement instead of an external product: embed inside a team's real workflow, diagnose the actual constraint, and ship the tool that removes it — then hand off so the team can run it without you.