Chen Ke-You

ResultsAnswered many quick-fire questions in class,Group ranked 2nd in class; produced two Canva decks and a three-minute short-talk video.Evidence ↗

Learned and didPeriodic table and semiconductor materials, diodes, LEDs, transistors, digital/analog conversion;I designed and tested the series/parallel circuits myself. The parts box I played with as a kid — burning out quite a few components — finally paid off in this course.

Final video assignment“How semiconductors affect human life” — my group chose Apple Vision Pro.The official record's reason: “it's hailed as a world-shaking masterpiece, but users say it has many flaws — so I chose it as my subject,Be a myth-buster” That was the first time I didn't take a single source as the conclusion.

In the classroom

ResultsMy class remarks were cited in three other groups' slides; Kaohsiung Education Expo curation, IExplaining transistors and chips to the public bilingually. In the official record I wrote: “Truly effective learning is being able to answer other people's questions.”Evidence ↗

Core theoryCPU vs. GPU architecture, Apple M vs. Intel, field-effect transistors, Planar FET to FinFET structural evolution, pros and cons. The intro course taught me what components are; this one taught me why they're shaped that way.

Hands-on makingnMOS switching experiments and Leakage-current testing, disassembling a Bluetooth speaker, Fritzing schematics, cost estimates; simulated in TinkerCAD then built on a breadboard: Interactive fan-speed control exhibit, the very one visitors turned by hand at the Education Expo booth.

The question that hit me hardestWhy phones overheat.The answer was the leakage current I measured myself — a transistor theoretically off still leaks current; multiplied by billions of transistors in a chip, that's the temperature in your palm.Textbooks don't write this — I measured it and connected it back to life before I truly understood it.FinFET wraps the channel with a 3D gate — solving exactly the same problem.

Industry insightthe CUDA platform, Intel/NVIDIA/TSMC revenue trends, and Analysis of NVIDIA's and Intel's declines and comebacks— analyzing not whose tech is stronger, but why a company misjudged and how it recovered. Another assignment was Writing to semiconductor professionals, that was when I first realized: what I want to know, I can just ask. Later writing to the website owner and designing interview guides were extensions of that move.

In the classroom

Hands-on work done:The assigned theme was climate-adaptation and net-zero energy sustainability; used the Golden Circle to think why before what.

Concrete results:Produced the smart grid capstone; and formed my view of AI — the teacher's words: “garbage in, garbage out; but throw gold in, and maybe great gold comes out.”

Hands-on work done:Learned engineering drafting and built the geometric model of a Strandbeest linkage mechanism in GeoGebra, verifying proportions and angles segment by segment.

Concrete results:Completed the full linkage geometry of the bionic mechanism, with basic competency in mechanism design.

GeoGebra workspace

Reflection:Wrong steps or wrong parameter settings both lead to failure. I made alignment mistakes many times; repeated practice has fixed that.

Evidence:GeoGebra original file (interactive model) ↗

Hands-on work done:Learned engineering drafting; the teacher scheduled exams to check our learning, with questions drawn from the Class-C draftsman license exams.

Setbacks:If I can't understand the question, drafting suffers. I need to break a question down first; otherwise the order of steps easily goes wrong.

My own drawings (Class-C draftsman license exam questions)

Reflection:I resisted at first, but later found a strong interest in and focus on engineering graphics.

Evidence (two Onshape documents):Onshape document 1 ↗, Onshape document 2 ↗

Hands-on work done:Experienced laser-cut drafting, from drawing the line art on a computer to fabricating the finished piece.

Reflection:The final step of removing the white border tests patience and care; after repeated practice I am fluent at it.

Inkscape computer drafting and laser-cut result

Six sites, each showing something different:

ASE GroupSeeing packaged chips, including a demo of Apple M2 packaging. My first time knowing a chip still has a long road after it's done.

STUST cleanroomEarly experience: gowning up, the air shower before entering. First time in a cleanroom suit, seeing the textbook wafer in person.Evidence ↗

NCKU Semiconductor Research CenterSaw equipment using TSMC's CoWoS packaging, testing instruments, andWhat students here research. It was the first time I realized the same topic draws different questions in a company than in a lab.

MacronixA sentence I heard inside a semiconductor company and still remember:Only once you've internalized it can you explain it.Later I relied on this in science-fair defenses and explaining at the Education Expo booth.

SEMI international semiconductor forum (SEMICON Taiwan 2024)Listening to experts on industry trends. My first time on the semiconductor industry's show floor — and my first time realizing that those textbook chapters are an entire industry out there.

“The World of Semiconductors” at the National Museum of Natural Science, TaichungThe grade-10 Advanced Semiconductors trip put the FET structures from class back into an exhibit context.Evidence ↗

Two industry engineers:One engineer each semester — one from TSMC, one from ChipMOS. What they gave me wasn't technology, but how to work. I recorded four lines in the official original.

My takeaways:“My horizons widened — I don't live in my own ivory tower.” My motivation for going to these places was simple:Seeing machines you normally can't.But after six sites, what I saw was really one thing: chips with different functions and components from different processes all end up packaged, interconnected, and integrated into a system that runs reliably.“Integration” is itself an engineering discipline requiring solid training— that image has shaped how I view my own work ever since.

One more thing outside the tech column:The teacher led the visit in Chinese and English — my first time using English in a real situation, not a classroom role-play. Later I could explain transistors to the public bilingually at the Kaohsiung Education Expo; that started here.

Proposed the direction, led the team and assigned work, supervised progress, revised and produced the written report and poster, built the Node-RED site, implemented the circuit boards and Node-RED code, consolidated information.

Division-of-labor evidence:The official team division-of-labor table lists me (team lead) at 55%, with teammates Weng Yu-Jie and Hsu Yu-En contributing jointly.

Architecture: simulated inputs → Node-RED rules-based decision layer (dispatch priority ranked by social external cost) → MQTT → cloud logging and light-signal alerts; hardware shown at scale with Arduino and breadboards.

Two design judgments worth explaining: (1) I included real ramp-rate differences (coal 5.5 MW/min vs. hydro 33 MW/min), because dispatch feasibility depends on whether units can keep up; (2) the load-shedding order was fixed as industrial → commercial → residential,Residential always last — that's a value judgment, not a technical one。

System hardware and screens

Used the Impact Pathway Approach (IPA) — the bottom-up monetization method recognized by the EU NEEDS/ExternE projects and the OECD; four steps: emissions inventory, atmospheric dispersion modeling, dose-response functions, monetization.

Sources include the Taiwan Energy Journal's external-cost analysis of coal-fired health impacts and Academia Sinica RCEC estimation formulas.Every self-made table in the appendix is labeled “drawn by the researcher” with the original data source attached.Evidence ↗

Data sources and records

Earthquake, typhoon, and reservoir panels deliberately do not auto-control the grid, for three reasons: mismatched time scales, incomplete information early in a disaster, and “whether to preventively shed load” being a risk trade-off requiring human judgment.

It's the same judgment I later held firm on in HERMES: AI responsibility boundaries.

Code generation can go to AI, but “how the system should be architected” and “whether the dispatch logic matches the grid's physical reality” are human responsibilities.

The team called this principle “80/20 Vibe Coding” in official documents; the ratio describes effort allocation, not a quantified metric.

The initial direction was “two-way distribution communication and load allocation” on the consumer side. The advisor challenged: “renewables fluctuate constantly — you can't control the generation side.” In team discussions I also said, “two-way communication is a dead end; I can't figure out how to do it.”

Finally, in oneAn hour-long team meetingIt pivoted toward real-time dispatch on the generation side, with three decisions: positioning around a “decision brain”; flipping the traditional “ Thermal-first, renewables-support” logic; and making “transparent decision-making” the key differentiator.

In late development I led cross-debugging with AI, using a “plan review” mechanism: code given to multiple models for debugging simultaneously, then reviewed item by item by me. In week 11 we found PM10/PM2.5 metrics mixed — a fatal problem — and immediately fixed it and refactored the Node-RED formulas to PM2.5-specific social-cost coefficients.

This error wasn't caught by my eye — AI repeated cross-checking found it.This taught me that academic citation isn't just plugging data into formulas — you must clarify the definitions and limits behind the data.

When the project needed Taiwan's real-time generation data, I found a third-party site organizing Taipower's open data clearly and in real time. On behalf of the team I wrote to the site owner under my own name requesting iframe embedding, and volunteered four commitments in the letter:

• Display only, no modification of content
• No commercial use
• Clearly citing sources in project documentation and oral presentation
• Never treating the results as developed by our team

Approved the same day (2026-02-12).Evidence ↗

Problem scale:One thermal power plant's major overhaul involved 275 primary tags、1,717 isolation operations、15 sections working simultaneously, with paper tags scattered across 40+ tag cabinetsin. One missed entry could become a risk.

The “tag” (work permit / equipment isolation tag) process during a thermal plant's major overhaul originally ran on an Excel workbook full of VBA macros plus physical paper tags scattered across dozens of cabinets, spanning 10+ sections working simultaneously.

Taipower already had a tag system, but it only opened, queried, and generated serial numbers — no cross-unit real-time collaboration.

What these three things share isn't technology — it'sWhere knowledge lives. The process ran on judgments held in senior staff's heads; newcomers could only learn by shadowing and word of mouth. What I did was move those judgments from heads into a system — searchable, traceable, handover-able. I later learned this has a name: Knowledge externalization。

The field side is an e-paper tag (Spectra 6, 4.3″) with QR-scan reporting; the backend is a web service with access control and a database.

The technical stack: a backend with SQLite serving a browser SPA; qrcode[pil] for QR generation, an xlsm→JSON import script, packaged as a Windows desktop app.

• The whole project de-identified externally: site, unit numbers, staff names, and equipment codes all replaced with fictional values
• Also created a separate “Security Audit Report” as the single source of truth, and wrote an idempotent cleanup tool
• Demo accounts created by the system, passwords hashed with werkzeug; in production mode, startup is refused without SECRET_KEY
• Audit technique findings:The first three audits found no sensitive strings — the root cause was that Chinese in .xlsm is stored as XML numeric entities, so direct Chinese matching always returned zero. The correct approach is to decode first, then match. This round of cleaning 2,433 site

Replaced single-use paper tags with e-paper for resource circularity; and made a magnitude-level estimate (noted in the deck as “magnitude estimate, not precise”) for the causal chain: tag errors → equipment misoperation → unplanned unit shutdown → extra fuel for replacement load → higher emissions.

Expansion planning:Single plant → replicate to sibling plants within the same company in six months with an NFC pilot → extend to thermal and private IPP plants within a year.

For the competition rules and five technical claims, I verified each against official primary sources and proactively overturned or corrected three of my own statements: a wrong belief about scoring weights, one e-paper use case with no traceable original news source, and one spec number sourced from a third-party reseller page rather than the manufacturer.

Three layers of problem awareness: information asymmetry, digital inequality for disadvantaged groups, and resource waste from mismatched demand (violating SDG 12). The problem came from real observation — salespeople in a group-chat screenshot who “couldn't even explain the processor” yet “enthusiastically recommended.”

The countermeasure wasA “spec translation system”Not a price-comparison tool: turning 120Hz into “smooth scrolling,” so non-tech people can search in everyday language like “clear photos of my grandkids” or “no lag in games.”

System screens

Prof. Tsai Meng-Fang's “don't be so scattered — focus on one point and take it to the extreme” was this work's most critical external feedback. Plus review comments from industry professionals on site and the head of Kaohsiung Industrial High School's computer science department.

Certificates and external review

Before submission, a Google policy change removed the old model and the system briefly couldn't connect to the API; I presented the results with a screen-recorded video instead.That was my first realization that technology selection must consider long-term access, not just what works now.

What I did:Explaining transistor principles to the general public, in Chinese and English.

What I learned:Explaining the same principle in two languages forces me to confirm I truly understand it.

What I did:The booth theme “Exploring Digital & Analog” — explaining that digital is “selective sampling” and analog is “continuous existence,” and using a high-end DAC chip to let listeners hear the gap between lost detail and true reproduction.

What happened:I had assumed vinyl worked via up-and-down grooves, and was corrected on the spot by Prof. Chen Chiu-Min of Soochow University's Physics Department — vinyl works via lateral displacement, and stereo comes from the difference in induced current between the coils on both sides of the stylus. He even wrote out the derivation on the spot.

On site that day

What I did:inKaohsiung Education Expo self-directed learning annual meeting (Youth Talks)Presented the phone-selection system in a speech-based competition format, to Bureau of Education judges and the audience.

What I learned:Turning a technical project into a lesson anyone can follow.Evidence ↗

I developed a methodology kit (53 files, 12 version iterations) that turns “one-sentence customer needs” into spec documents and go-to-market plans, by repeatedly having AI build while I gatekept.

What's really worth telling is the hard rule I set inside: AI may never fill the “actual results” field; with no data, KPI fields stay blank — no guessing from impressions. I wrote at the time:“‘Pending verification’ in practice isn't honesty — it's pre-emptively pushing responsibility into the future.”

Define the role, explain the task, set constraints, specify the output format.

First tell an AI the requirements and ask it to draft the prompt; after I review and revise it, hand it to the AI doing the actual work.The industry calls this meta-prompting; I don't claim originality.

Before submitting, I asked AI to play “the strictest reviewer” and tear its own project apart. I even created a file named “find-trouble” and had it return a critique report with serious flaws and data-credibility challenges.

Treating implementable AI as the most precious resource, reserving it for tasks that truly need a system built; plain search, fact-checking, and first-pass organizing all went to the web version.