Hengzhen Jiang

Master's student in Financial Technology at Duke University. I've worked in user research, data analysis, and building practical AI workflows.

See my work Get in touch

↓Scroll

Timeline

Education

AUG 2025 — JUN 2027In progress

Duke University

Master of Engineering in Financial Technology

DURHAM, NORTH CAROLINA

Coursework: Machine Learning, Statistical Analysis, Asset Pricing and Risk Management, Python Programming, Data Processing and Algorithms.

SEP 2021 — JUN 2025Completed

Shanghai Lixin University of Accounting and Finance

Bachelor of Economics in Financial Engineering

SHANGHAI, CHINA

Honors: Shanghai Outstanding Graduate; Outstanding Student Leader.

2021 — 2027 · Most recent first

2026
Jul — Aug 2026
Jul — Aug 2026Multimodal Short-Video Content Analysis AgentSide project: a workflow for reviewing creator videosProject
Jun — Aug 2026
Jun — Aug 2026Lovart · International KOL Operations InternCreator campaigns across Europe and North America, plus a LangChain workflowInternship
Jan — Apr 2026
Jan — Apr 2026Ethereum Token Fraud Detection27 behavioral features from 2.2M on-chain transactions, LightGBM at 0.96 AUCProject
Jan — Apr 2026
Jan — Apr 2026iQIYI International · North America Marketing Operations Intern33,876-response annual user study across English-speaking marketsInternship
2025
Aug 2025 → 2027
Aug 2025 → Jun 2027Duke University · MEng, Financial TechnologyMachine learning, statistics, asset pricing, and data engineering courseworkEducation
Feb — May 2025
Feb — May 2025Dewu · Business Analysis Intern, Beauty Card ProgramSegmentation, A/B-tested outreach, and a live activation dashboardInternship
2024
May — Aug 2024
May — Aug 2024SDIC Securities · Industry Research Intern, HealthcareMarket sizing and a three-scenario valuation model for a 10,000-word reportInternship
2021
Sep 2021 — Jun 2025
Sep 2021 — Jun 2025Shanghai Lixin University · BEcon, Financial EngineeringShanghai Outstanding Graduate · Outstanding Student LeaderEducation

Skills

Tools I use

Data & analytics

SQL, Python, Excel, Tableau, Power BI

Pandas, NumPy, scikit-learn, and Matplotlib in Python; VBA and PivotTables in Excel; A/B testing on live campaigns.

Modeling & interpretation

LightGBM, SHAP, regression, DCF

Five-fold cross-validation, feature attribution, and three-scenario valuation models with sensitivity analysis.

AI & automation

LangChain, agent workflows, Codex

Prompt engineering, multimodal content analysis, and API integrations for internal tools.

Languages & communication

Mandarin (native), English (TOEFL 105)

Research reports, content planning, and creator outreach in both Chinese and English.

Work experience

Four internships between 2024 and 2026.

1.0 — Lovart · Jun — Aug 2026

Creator campaigns in the U.S. and Europe, plus one automation workflow

International KOL Operations Intern · Beijing Qidian Xingyu Technology · Global growth and AIGC product marketing

Lovart is an AI design and video tool that was expanding into Western markets. I worked on its creator campaigns in the U.S. and Europe: campaign positioning, creator outreach and content review, and a small workflow that kept our collaboration records in sync.

1.1Campaign strategy

Planned the Lovart × SMB and Lovart × Seedance 2.5 campaigns for Europe and North America. Analyzed competitors and creator segments to position Lovart's brand-design and AI-video capabilities for SMB owners, designers, and video creators.

1.2Creator operations

Managed creator partnerships across TikTok, Instagram, and YouTube, reaching 500+ creators. Delivered 85 sponsored posts with 10M+ total impressions while keeping Instagram and YouTube CPMs near $15 and $100.

1.3Content quality

Reviewed product demonstrations, messaging, and creative quality against product-understanding and expression standards, with feedback rounds before publishing.

1.4Workflow automation

Built a LangChain-based KOL information-processing agent and integrated Python with Feishu APIs to synchronize collaboration records, deduplicate matches, and backfill publishing data.

Campaign strategyLangChainPythonFeishu API

Collaboration record syncLANGCHAIN → FEISHU

Creators reached

500+

US · EU

Record	Action	Status
Design-segment creator · YouTube	Backfill publish date	Synced
SMB-segment creator · TikTok	Match duplicate entry	Merged
Video-creator segment · Instagram	Sync collaboration terms	Synced
New inbound · form submission	Create record	Running

What English-speaking viewers watch, and how they find it

North America Marketing Operations Intern · Audience research

iQIYI International runs an annual user survey across English-speaking markets. I handled the content analysis: 33,876 completed responses, analyzed in Python by content preference, membership status, age, country, and where people first discovered the content.

2.1Content preference

Compared awareness, viewing interest, and preference across Asian dramas, short-form dramas, animation, and AI content. Content does not play a single role in growth: Chinese dramas form a consumption base, BL/GL resonates with younger members, Korean dramas and animation show stronger acquisition potential.

2.2Audience segmentation

Cross-analyzed country, age, membership status, and ethnicity to find where the same content carries different value.

2.3Growth pathways

Short-form drama is discovered more often through social media. For non-users, social channels are the opening to create discovery and guide interest toward iQIYI.

2.4Output

A research report informing English-region content selection, off-platform discovery, and product consideration.

PythonSurvey designSegmentationVisualization

English-region annual studyiQIYI INTERNATIONAL

Completed responses

33,876

cleaned

Content directions

compared

Analysis layers

linked

Asian dramasShort-form dramasAnimationAI content

2.1

Content preferenceWhat role does each direction play in English-region markets?

2.2

Audience segmentationCountry · age · membership status · ethnicity

2.3

Growth pathwaysWhere people discover content and how they enter iQIYI

Different content categories serve different moments in the user journey.

3.0 — Dewu · Feb — May 2025

From behavioral data to better activation

Business Analysis Intern · Beauty Card Program · Commercial analytics

At Dewu I supported the beauty-card membership program. I used SQL and Python to segment users, helped design the SMS outreach and its A/B test, and built the dashboard we used to track activation and subsidy spend.

3.1Segmentation

Queried and cleaned behavioral data with SQL and Python, then separated high-intent and broader-intent audiences by purchase frequency and brand-browsing depth.

3.2Outreach and testing

Designed differentiated SMS outreach and validated it through A/B testing: 500+ new card activations and roughly a 12% month-over-month lift in activation rate.

3.3Merchant selection

Built a multi-criteria Excel model to select campaign merchants, using placement exposure as the incentive for merchant discounts.

3.4Monitoring

Built a BI dashboard for activation and SKU-redemption funnels, shifting placement allocation from static ranking to data-driven iteration and reducing subsidy costs by about 15%.

SQLPandasA/B testingExcelBI

Beauty card · activation funnelDEWU · COMMERCIAL ANALYTICS

New cardholders

500+

contributed

Activation lift

+12%

month over month

Subsidy cost

−15%

approx.

A · generic SMS

baseline

B · segmented SMS

+12% activation

Funnel widths are schematic. Reported figures are the ones above.

4.0 — SDIC Securities · May — Aug 2024

Sizing two ophthalmic markets and valuing one company

Industry Research Intern · Healthcare research group

In SDIC Securities' healthcare research group, I estimated the size of the intraocular-lens and orthokeratology-lens markets from Choice terminal data and penetration benchmarks, then built the valuation model behind a 10,000-word report on Aibo Medical.

4.1Market sizing

A top-down comparison of domestic and overseas penetration rates showed both markets early in adoption, with a credible domestic-substitution case. Regression forecasts of 2026 market size kept error within ±8%.

4.2Valuation

Broke down Aibo Medical's segment revenue and margin trends, set growth assumptions by competitive position, and built a three-scenario DCF plus P/E model with sensitivity analysis on the key variables.

4.3Output

The analysis went into a 10,000-word research report with an investment recommendation.

Choice terminalExcelRegressionDCF

Market sizing & valuationSDIC · HEALTHCARE

Forecast error

±8%

2026 market size

DCF scenarios

plus P/E

Report length

10,000

words

4.1

Penetration benchmarksIntraocular lens · orthokeratology lens · domestic vs overseas

4.2

Regression forecast2026 market size for both tracks, error within ±8%

4.3

Three-scenario DCF + P/EAibo Medical segment revenue, margins, and sensitivity

Sources: Choice terminal financial and industry data.

Projects

Two projects, one from a course and one on my own.

5.0 — Academic project · Jan — Apr 2026

Spotting spam tokens from on-chain behavior

Ethereum Token Fraud Detection · Team research project, Duke University

A team course project at Duke. For Ethereum tokens that imitate the symbols of verified assets, we built a classifier from on-chain behavior, using roughly 2.2M transfer events to compare patterns across token contracts.

5.1Key finding

Spam tokens cluster in a narrow block window and show concentrated distribution. 34.0% of spam tokens were active in only one block, against 7.2% of legitimate tokens.

5.2Method

Engineered 27 behavioral features over 3,606 labeled tokens and benchmarked six classifiers. Symbol collisions were used to construct labels, not as model inputs, so the classifier learned on-chain behavior rather than the labeling rule. LightGBM tuned with five-fold cross-validation reached 0.89 macro F1, 0.96 AUC, and a 90% fraud-detection rate.

5.3What did not work

Semi-supervised expansion with about 8,750 unlabeled tokens did not improve performance. Distribution shift and pseudo-label noise made the supervised model more reliable.

5.4From model to rules

Applied SHAP to surface fraud signals such as short active windows and concentrated senders, and turned them into risk-control rules and a cost-aware decision threshold. I worked across the pipeline: framing the problem, building labels, feature engineering, model comparison, and interpretation.

PythonLightGBMSHAPscikit-learn

Global feature importanceMEAN |SHAP| · 541 TEST TOKENS

Macro F1

0.8838

held-out

ROC–AUC

0.9562

held-out

Labeled tokens

3,606

27 features

n_distinct_blocks

0.81

block_range

0.75

value_mean

0.68

top1_sender_share

0.47

transfers_per_block

0.42

n_unique_senders

0.36

unique_values_count

0.35

n_unique_receivers

0.32

Rather than token names, the model relied most on activity spread, transfer value, and sender concentration.

6.0 — Independent project · Jul — Aug 2026

A short-video analysis agent for reviewing creator content

A side project from the Lovart internship. It collects public short videos, transcribes the subtitles, samples key frames, and uses an LLM to note the hook, where the product appears, whether a sponsorship is disclosed, and the call to action, so reviewing creator content takes less time.

Collects public short-form contentTranscribes subtitles and reads key framesFlags hooks, product exposure, sponsorships, and CTAs

Agent template

Name

Content Analysis Agent

Scope

Overseas AIGC growth · creator-content review

Instructions

Collect public short videos, transcribe, sample key frames, and report hooks, product exposure, sponsorship signals, and calls to action.