Building things that make sense to the people who actually use them.
Trying to think clearly when most things compete for attention.
Quietly obsessed with details most people skip.
[BUILDER][CREATOR-LITERATE][AI-NATIVE]
joanduan.dev · v0.3
NOW · ingest · learn · ship
·perception
·comprehension
·composition
– ABOUT –
Probare et Aedificare.
To prove, and to build.
THE MOVE
Five cities, three languages, two school systems before I turned 18 — Shenzhen, London, Paris, Frankfurt, Munich.
Mandarin at home, English across continents, German learned by immersion. I learned to read environments faster than I learned grammar. That's still how I work.
THE VOICE
I grew up impulsive — quick to opinion, slow to proof. Then I moved to a country where every sentence has to carry its own evidence. It rewires you. Mine got rewired.
Now: I notice patterns first, name them last. Quality travels further than volume. The best ideas in a room are the ones nobody's made simple yet — and made true.
THE CLIMB
I keep picking the harder track. High school, university, every choice between an easier path and a steeper one — I go steeper. Not for the view. For the legs you build climbing it.
Here I'm surrounded by people years ahead of me on the curve. I'm not intimidated — I'm calibrating. The point of climbing isn't to be the tallest. It's to keep finding harder mountains.
THE PIVOT
I chose Wirtschaftsinformatik because the bridge between business analysis and code is where the real problems live. Pure tech misses the why. Pure business misses the how.
I'd rather build the bridge than guard one shore.
THE BRIDGE
My mother is an engineering designer. She designs hardware — conferencing systems and the equipment deployed in meeting rooms around the world.
I grew up watching her work. Drawing on the screen, running the meeting, making the call. Every product she shipped did two things: contributed to the business, and got used by real people in real rooms.
She does this in mechanical engineering. I want to do it in computer science. Different stack, same kind of work — sustainable products that create value for the business and the people who use it.
CURIOSITIES
Things I'm exploring beyond the day-to-day: product strategy, human-computer interaction, applied AI in everyday tools. Different pace, different scale — but every one of them is on its way to becoming work, not staying as wishlist.
THE BELIEF
The best technology is the kind you forget you're using — because it works the way you already think. The worst makes you translate yourself before it understands you.
The work, for me, is closing that gap.
TUM · CIT
Haichen Duan
Designer · Engineer · Analyst
B.Sc. Information Systems
Double-tap to flip
TUM · WIRTSCHAFTSINFORMATIK
Contact
Haichen Duan
Location
Munich, DE
Probare et Aedificare.
– EXPERIENCE –
Where the resume lives.
2026 – Present
Intern, Quality Management Digitalization
BMW Group · Munich
—Own the end-to-end cleansing of large, heterogeneous production reporting data: identify erroneous and inconsistent records, trace their root causes (including faulty calculation logic in the legacy solution), correct them and migrate the clean data from SharePoint into the internal platform.
—Transformed a grown, rule-less Excel solution that produced incorrect results as data volume increased into a fixed, standardized data format with defined validation rules, making the internal platform scalable and maintainable in the long term.
—Built a digital workflow on the platform so stakeholder meetings work from one consistent, validated data basis instead of manually compiled Excel reports, making these meetings more efficient.
—Developer and administrator responsible for the platform in production use: clarify requirements directly with management, order coordination and shop-floor users, translate between business and technical language, and analyze and fix issues reported by users.
—Currently building a barcode scanning tool for 100% sorting inspections that replaces manual Excel capture (database layer with PL/SQL business logic on the test environment, frontend prototype with handheld scanner integration).
An agentic AI repository auditor. Paste any GitHub URL, get a structured 5-dimensional audit (documentation, architecture, maintenance, testing, security) in under a minute, plus decision memos, dependency due diligence, and side-by-side repo comparison.
IVI Defect Triage: Does Symptom Normalization Improve Duplicate Detection?
A controlled ablation study on whether normalizing noisy defect reports into one-sentence symptom summaries improves embedding-based duplicate detection, extended into a full triage pipeline and a tool-use agent.
A context-aware re-finding web app for tourists — surfacing places you saved at the moment they become relevant. Sole frontend on a 2-person team. Built four paper-faithful features that translate a 2017 academic recommendation framework into UI signals users can actually read.
A Python CLI agent that parses unstructured sales chat logs into schema-compliant Salesforce payloads in under five seconds.
PythonClaude APISalesforce APISQLite
2026.06Planned
○ROADMAP
DATA / FORECASTING
Industrial Sales Forecasting
A hybrid SARIMAX/ETS modeling system to identify growth potential and customer lifecycle patterns across B2B industrial sales accounts.
PythonPandasStatsmodelsScikit-Learn
2026.06Planned
○ROADMAP
AI / RESEARCH
AI Vision Evolution
A side-by-side comparison of how vision models read the same image — from early CNNs through Vision Transformers to multimodal LLMs. A study in how machines have learned to see.
PyTorchGrad-CAMClaude Vision
2026.06Planned
○ROADMAP
AI / NLP
NLP on Creator Comments
Sentiment and theme extraction across tens of thousands of audience comments. Listening at a scale no creator can do by hand.
PythonLLMspaCy
2026.07Planned
– DATA LAB –
Smaller experiments.
Focused analytical notebooks and dashboards — narrower in scope than the projects above, but built end-to-end on real data.
DATA / ANALYTICS
● LIVE
P&C Insurance Analytics Dashboard
An interactive 4-view dashboard on French Motor TPL (Charpentier 2014) — surfacing where a €234M-exposure portfolio actually loses money: 6 unprofitable regions, the 18-25 driver cohort, and a healthy BonusMalus retention curve.