About

From atoms to tokens: map the supply chain, find the bottleneck, build the loop. #

Bridging atom factories and token factories. Atom factories build AI’s hardware — chips, servers, power. Token factories (data centers) turn that hardware into intelligence. My work lives in the translation between the two: between the physical constraints of supply, and the AI workloads that run on top. I’ve worked both ends, and both sides of the supplier table.

Product = Analysis + Process.
Analysis maps the supply chain and finds the bottleneck. Process turns the answer into a repeatable loop that can scale beyond one person.

Hi, I’m Chung-Hao (CH) Lee 李崇豪 — based in Sydney.

Since 2018, my work has followed one thread: find the physical single points of failure behind complex systems, then build the systems that manage them.

I started on factory floors and grew into AI systems. At Pegatron Suzhou I led a 15-person cross-functional team operating the Google Home manufacturing line — around 2M AIoT units a year — and learned what a supply agreement is worth when a line is down and the supplier isn’t yours.

As a Supply Chain Analyst at the University of Maryland, I owned the semiconductor portion of a National Defense Industrial Association study, mapping raw-material, supplier and logistics risk and the single points of failure inside them, using public data only. The work was covered by National Defense Magazine.

At Wistron Taipei — a global Tier-1 ODM for Dell — I brought industrial AI to Dell laptop manufacturing, delivering $1.2M in annual cost savings through AI-driven predictive maintenance across 20B+ telemetry records scaling to 40B+. Dell’s SVP of Worldwide Procurement signed the 2023 award for the programme I led. Sitting on the supplier side taught me the other half of the table: how a quote is built, where the margin sits, and what a customer can and cannot be asked to pay for.

Now at TikTok Sydney, I point the same method at AI itself. I built TikTok Live’s Root Cause Analysis system — process and platform — from 0 to 1, tracing why content-moderation models fail and routing each fix to the team that owns it. Its third generation, RCA Copilot, is an AI agent where humans and agents work in one system.

Two years ago I started reading down the stack: from the models I operate, to the servers under them, to the power and land under those. That reading became the Data Center series, an Australian siting framework, and a daily intelligence feed I run for myself. It’s the direction I’m heading — the buying side of data centre capacity: what it costs, where it goes, and what the deal says.

The through-line isn’t a domain. It’s the method: map the supply chain, find the bottleneck, then design repeatable workflows that close the loop.

Both Sides of the Supplier Table #

Most people who buy have only ever bought. I’ve done both, and the second half is what makes the first half work.

Contract manufacturing runs on a two-sided relationship. At Pegatron I was the factory — holding the line, chasing suppliers, carrying the cost when something upstream failed. At Wistron I was the supplier — building the quote, defending the price, asking a Tier-1 customer to pay for something. The two jobs read the same contract from opposite ends, and only one of them teaches you where a quote is soft, when a counterparty genuinely cannot deliver, and when they are simply holding a position.

Buying capacity is that same job at a different scale: securing supply from a counterparty you don’t control, at a price you can defend, under a contract that decides who carries the risk.

Data Centres — What I’ve Built and Published #

For the past two years I’ve been working down into the commercial side of data centres: how capacity is priced, where it gets built, how it’s financed, and what the contracts say. Self-directed, and published as I go.

Data Center 101 — a self-taught 14-part series working down the stack. TCO economics and why this is an OPEX-dominated business, power, cooling, the supply chain from copper mines to AI clusters, reliability tiers, DCIM, and a Sydney regional deep-dive.

A data-driven siting framework for Australia — data centres have to go somewhere, and there was no public, evidence-based way to argue about where. My GovHack 2025 entry built one: every major Australian region scored across five factors — power, cooling, connectivity, talent, land — using AEMO grid data, Bureau of Meteorology temperature and water data, Telstra InfraCo fibre and submarine-cable maps, Tech Council workforce data, and ABS population data. It put Brisbane and the ACT first, and scored Sydney’s power zero. A year on, grid capacity is the constraint this market talks about most, and operators are being pushed toward Brisbane.

My Daily Digest — keeping a thesis current is a data problem, so I automated it. A daily brief I built and run for myself across 13 data-centre sources, including nine Australian operators and the AEMC and AEMO grid feeds, plus geopolitical risk and AI-builder tracking. GitHub Actions and free LLM APIs, delivered to Telegram in Traditional Chinese.

Capital and cycles — infrastructure is financed long before it is proven, so I write about the money as much as the machines: Why Global Capital Is Betting on Sydney’s AI Data Centers · The AI Supply Chain Dividend · From the Coolidge Prosperity to the AI Boom · The Global AI Training Pivot.

Experience #

TikTok — Product Operation Manager, Model Operation (AI/ML) #

Sydney · 2025 – present

I own Root Cause Analysis for TikTok Live’s content-moderation AI — the process, the platform, the models that evaluate it, and the day-to-day operation that keeps production models improving.

RCA process + platform — built both from 0→1 #

Designed the end-to-end RCA process from scratch: how issues surface, how the right cases are found, how root cause is analysed, and how each finding is routed to the owning team — algorithm, labeling base, engineering, or policy — for the fix.
Built the platform that operationalizes that process: consolidated data scattered across disconnected systems into one standardized workspace, unifying discovery, investigation, analysis, and issue hand-off in one place. Authored 15+ PRDs across Engineering, Data, Ops, and Policy.
Cut end-to-end RCA time ~56% across the core team; 40+ active users, 10,000+ monthly queries, adoption spreading beyond RCA to adjacent functions.
Third generation, RCA Copilot: an AI agent that plans the evidence it needs, retrieves it through purpose-built skills, and traces every model error to a single root cause — humans steering via human-in-the-loop. Evolved from human checklists → decision-tree attribution → agent platform.

AI model building + evaluation pipeline — from 0→1 #

Built AI models from scratch to clean and validate data at scale (prompt engineering, multi-agent orchestration, agent-looping), producing the high-quality datasets used to evaluate production models.
Stood up the team’s first systematic, repeatable model-evaluation pipeline: 1M+ cases across 62 iterations at ~99% labeling accuracy over 13 batches.

RCA operation — 35 live models #

In the system’s early days, ran RCA across all 35 production models moderating live content (nudity, hate speech, bullying, and more) — broad coverage while the process was still forming.
As the RCA platform matured and the operation scaled across the organisation, my scope focused to owning nudity detection end-to-end in greater depth — raising that model’s F1 score 38% (to 80%+) through multi-agent orchestration with an arbiter agent and describe-then-verify hallucination control.
Established a recurring Product–Ops–Policy review that turns root-cause findings into owned fixes — a standing governance forum with named accountability, replacing case-by-case escalation with a systematic feedback loop from risk discovery to downstream ownership.

What this side of the table taught me #

I operate the workload that data centres exist to serve. Running 35 production models day to day is a direct education in what compute demand actually looks like from the inside — how it grows, where it spikes, and what happens when the capacity underneath it is wrong.

TikTok Spot Bonus Award, 2025 Q4 — for independently leading the RCA platform with strong ownership and cross-functional delivery.

Wistron — AI Project Manager #

Taipei · 2022 – 2024

Worked inside a global Tier-1 ODM for Dell — the supplier side of the hardware supply chain.
Led a three-phase industrial AI program for Dell laptop manufacturing (smart diagnosis → predictive forecasting → knowledge consolidation) across 6 cross-functional teams and 20B+ telemetry records, scaling to 40B+; also deployed LLMs for log analysis at scale.
Results: 70% faster debugging · 50% fewer return-to-repairs · $1.2M annual savings.

The goal was never to win the argument. It was to keep the programme funded.

The programme ran on telemetry streaming off machines in the field, stored in the cloud on our side of the relationship. It worked well enough on laptops that the customer wanted desktops added — several times the data, and more storage than we were paying for. Nothing further could run until someone bought the capacity, and nobody had budgeted for it.

I sized the growth and chose the option that fitted the existing stack over the cheapest one, on the view that migrating later would cost more than the difference now. Then I put the number to the customer. Their opening position was that the storage sat on our infrastructure, so it was our cost to carry.

I did not argue about whose infrastructure it was. Arguing over ownership only produces two versions of the same story. I rebuilt the case from the record — why the expansion was needed, why this option and not the cheaper one, how the price was derived — and proposed splitting it. That reframing was the whole move: from whose cost is this to how do we share it.

We settled on a shared cost. Neither side won, which is exactly why the next expansion started from a precedent instead of a fight.

Pegatron — Factory Project Manager #

Suzhou · 2018 – 2020

Led a 15-person cross-functional team operating the Google Home manufacturing line — 2M AIoT units per year.
Owned end-to-end SMT + FATP process control on the factory floor.

The goal was to get the line running again. Replacing the supplier was never the way to do it.

An optical module on a line I ran began failing the customer’s cosmetic inspection, and we stopped shipping. Fixed cost accrued every day the line stood still.

The module supplier had been nominated by the customer and held no contract with us: no audit right, no information right, no obligation to explain itself. Our own line checked clean within a week, which put the cause inside a company I had no standing to compel.

A second source could not carry the volume in time. Whatever happened, the incumbent had to come back and supply properly, which meant what I needed was not a substitute — it was something that would make him move. I had nothing that could.

So I built one. I qualified a second-source module for real, alongside the customer’s supply-chain team, and kept pressing for the process records. The qualification itself was the argument, and it never had to be said out loud. The change history came out: the supplier had switched adhesive vendors a month earlier, and the new adhesive outgassed during vacuum assembly and condensed inside the lens. He corrected it, supply stabilised, and the line restarted.

When you have nothing that makes a counterparty move, build it. Leverage is manufactured before the conversation, not discovered during it.

Recovering the cost, and choosing what to give up #

The second goal was not full recovery. It was a bill the customer would agree to share.

The stoppage left a real loss, and because the supplier was the customer’s nomination, we put the claim to them. They pushed back: the fault was not wholly theirs.

I did not contest that. Fighting over whose fault it was would have gone nowhere, so I moved the question from whether to how much, and broke the loss into a line-by-line bill — materials, rework, overhead, idle direct labour — each item with its basis and its evidence. In a money conversation what persuades is never grievance. It is an account the other side can audit and cannot fault.

Then I discounted the idle-labour line myself. It was defensible, because those operators could be redeployed and the full figure was never truly lost. It was also a signal: we were not there to squeeze the last dollar out of the account.

Giving up one line bought something worth more than the line — a counterparty still willing to deal with us the following year. The aim is not to win the round. It is to make sure there is another one.

University of Maryland — Supply Chain Analyst & Research Assistant #

College Park · 2020 – 2022

Owned the semiconductor portion of an NDIA study on supply-chain risk in defence-critical equipment — mapping single points of failure using public data only; covered by National Defense Magazine.
Research became the Springer-published MiLB → MLB promotion prediction model, featured in Maryland Today.

Earlier: Substitute Military Service at Taoyuan International Airport (2016 – 2017) — honorably discharged with a Ministry of the Interior certificate of merit.

Education #

MBA + MS Information Systems — University of Maryland, College Park (2020 – 2022), Smith Fellow Scholarship MBA major in Supply Chain Management; MS in Information Systems with a machine-learning focus — the two halves I still use together.

BS + MS — National Tsing Hua University, Taiwan

Selected Achievements #

TikTok Spot Bonus Award, 2025 Q4 — for leading RCA Copilot 0→1
Dell Outstanding Leadership & Execution Award, 2023 — presented to the Wistron AI Team for the programme I led, signed by Dell’s SVP of Worldwide Procurement
Wistron Excellent Digital Transformation Project Award, 2023
Springer-published research — Baseball Informatics: From MiLB to MLB Debut; featured in Maryland Today · UMD Outstanding Graduate Student Award nomination, 2022
Impact Consulting Fellowship, 1st Place (2021) · Hearst Data Competition, 3rd Place (2022) · Graduate Research Day, 2nd Place (2022)
VP, Smith Masters Supply Chain Management Association & APICS

Earlier: Scrum Master certified (2023) · NTHU Dean’s List ×3

Also Built in Public #

I tend to solve my own problems by building something, then open-sourcing it.

multi-lens-thinking — a six-stage AI reasoning pipeline that routes a question through four parallel analytical lenses before answering.

WikiPolyDraft — an open-source tool for AI-drafted Wikipedia translations. First article: Barangaroo, Sydney.

I also write about how humans and AI should divide the work — Beyond the Interface: Designing a Common Language for Humans and AI. Most people are still using AI without having decided what they think about it; that essay is what I think.

Let’s Talk #

If you’re working on the commercial and physical side of AI infrastructure — data centre capacity, siting and power, procurement and contracts, or the supply chains underneath — especially in Sydney or across APAC, I’d genuinely enjoy comparing notes. I’m equally happy being told where my reading of the market is wrong.

Reach me by email or on LinkedIn.