Max Zhang | Product, Data & Creative Practice

The product principle: automate repetitive healthcare work without hiding consequential decisions from the people responsible for them.

Case 01 / Expert configuration

The bottleneck was translation, not Excel.

Laboratory specialists defined biomarkers, ranges, thresholds, categories, and recommendations in spreadsheets. Engineering then interpreted the same rules and entered them into the product by hand. Every revision created another handoff.

I owned product definition and cross-functional delivery with laboratory, engineering, design, data, and operations partners. I chose a dual-path system: a structured editor for recurring work, plus an upload parser that preserved the laboratory team's familiar authoring workflow. Engineering approval remained a required state.

Observed operational result: setup moved from approximately one month to one week, an estimated 75% shorter cycle.
System effect: engineers shifted from repeated entry to review, approval, and exception handling.

Case 02 / Provider workflows

Use AI for coordination, not clinical judgment.

Provider interviews pointed to the same pattern in small practices: clinicians were spending time on intake, reminders, refills, follow-ups, and status checks because they had little administrative support.

I had feature-level ownership for an AI-assisted practice workspace. I designed a template library, persistent patient context, refill controls, and an updates inbox organized around Issues, Reports, Operations, and Archived. The useful role for AI was assembling a low-risk workflow from clear building blocks. In this reconstruction, I make control explicit by asking the provider to review the trigger, patient scope, and action.

Delivered scope: research, product definition, interaction design, and feature specifications.
Evidence limit: the product launched after I left, so I do not claim post-launch adoption or business impact.

Proposed evaluation

Trust is measurable at the point of control.

I would compare templates with manual setup using activation rate, completion rate, time to activation, provider edits before approval, disable or undo behavior, and escalations caused by ambiguous inputs. These are proposed measures, not historical results.

Company details are anonymized. The interface is a reconstructed portfolio prototype based on workflows I owned. The explicit review pattern shown here is a proposed production safeguard, not a claim about the shipped implementation.

Marker configuration

Draft panel

Foundational health

MarkerUnitRangeStatus

Vitamin D, 25-OHng/mL30-100Editing

Apolipoprotein Bmg/dL60-120Ready

Ferritinng/mL12-150Ready

Reference range30 to 100 ng/mL

Patient groupAll adults

DeficientInsufficientNormalHigh

Lab authored→Rules validated→Engineering approval

Try the provider workflow

Remind the patient 7 days before the prescription ends.

When7 days before refill

→

DoSend reminder

→

ThenCreate staff task

Draft needs provider review

Illustrative interaction. It demonstrates template selection, patient context, and provider confirmation.

The analytical decision: optimize for recommendations a marketing team could act on, not for a model that was impressive but difficult to explain.

Question and role

Who responds, and what should we send next?

The client's behavioral, email, and content data lived across more than 15 tables at different levels of detail. Within a five-person Fuqua team, I worked across data preparation, segmentation, interpretation, and presentation design.

Helped build a 2.66 million-row dataset with 76 engineered features.
Ran PCA and K-means, using elbow analysis to select four audience segments.
Led insight development for two segments and translated model output into decisions about audience, timing, and content.

Tradeoff and correction

Clarity mattered more than protecting our first approach.

I proposed a cluster-by-cluster narrative; the team chose an aspect-based structure after discussion. Later, a misunderstanding about the client's 240-day analysis window surfaced after modeling had begun. I returned to the client, clarified the requirement, and helped update the pipeline instead of defending an assumption.

Delivered findingCloser content matching did not always mean better response.

Past activity was the strongest predictor. Content matched too closely to a user's prior buying stage was associated with roughly three times lower response. The team delivered four audience segments with tailored email recommendations; we did not measure downstream campaign lift.

The segment names and visual are reconstructed for this portfolio. Dataset scale, methods, findings, and team scope reflect the client deliverable.

Audience model workbench

15+source tables

2.66Mbehavior rows

76features

38components

4audience clusters

Model viewDifferent models answered different questions

Decision treeExplain behavioral splitsInterpretable

Random forestRank predictive signalsRobust

Logistic regressionRead direction of associationDirectional

Past activity strongestEarlier week positiveOvermatching negative

Segmentation viewK-means after PCA

Behavioral intensityContent depth

Active explorersFocused evaluatorsQuiet researchersReady returners

Reconstructed analysis view. Segment labels are illustrative; dataset scale, methods, and findings reflect the team deliverable.

The growth problem: make an anti-racism mission easier to enter without making the issue feel smaller.

Role and decision

Warmth could invite action without softening the purpose.

As Marketing Manager at 1Thing, I shaped brand tone, color direction, audience positioning, and the strategy connecting the website, social content, and campaigns. Audience research supported a gentler, conversational voice instead of institutional or confrontational language.

I created “1 Thing Story,” a repeatable editorial format built around one person, one action, and one story. It gave volunteers a clear way to participate and gave the organization a consistent content system.

Built KPI reporting across social engagement, landing-page clicks, and completed campaign actions.
Observed an approximately 20% improvement in a blended index across comparable pre- and post-optimization periods.

Evidence limit

Useful direction, not a single-channel causal claim.

The 20% figure is a working composite of normalized engagement rate, landing-page click-through, and completed actions. It is directional evidence across comparable periods, not an audited lift attributable to one campaign.

Supporting growth proof

Consistency, collaboration, and community can compound.

Earlier, I independently produced gaming tutorials and livestreams on Bilibili, collaborated with other creators, and managed the community. The channel added approximately 30,000 net-new followers in three months.

The research and experiment board is a portfolio reconstruction. My 1Thing contribution covered brand tone, audience positioning, marketing strategy, and measurement, not the website's interface design or implementation.

Campaign learning board

Interview and audience synthesisPatterns carried into the campaign hypothesis

Specific actions felt easier to join than broad advocacy language.

Personal stories made the mission more relatable.

A conversational voice created more room for participation.

Research pattern→Message hypothesis→A/B comparison→KPI review

ControlInstitutional voice

Mission-first framing with a broad call to action.

vs

ChallengerConversational voice

One person, one action, and one clear next step.

Engagement rateAttention

Landing-page CTRIntent

Completed actionsParticipation

~20% directional composite improvement

The comparison shows the testing logic. The result is a working composite across comparable periods, not a single-channel causal lift.

MaxZhang

Three fields of work.Four case stories.

The bottleneck was translation, not Excel.

Use AI for coordination, not clinical judgment.

Trust is measurable at the point of control.

Who responds, and what should we send next?

Clarity mattered more than protecting our first approach.

Warmth could invite action without softening the purpose.

Useful direction, not a single-channel causal claim.

Consistency, collaboration, and community can compound.

I work where behavior, systems, data, and story overlap.

Moving images.Things made by hand.

Studio work