TikTok Mall & Shop Ads
$0M+
daily revenue reached by an ads delivery system built from 0 to 1, globally
Staff ML Engineer @ TikTok · Seattle, WA
Sitian Lu
I build ranking models retrieval engines ads delivery systems ML teams that turn marketplace signals into revenue.
Staff Machine Learning Engineer and tech lead at TikTok. Seven-plus years turning product, user, and marketplace signals into large-scale ML systems — and measurable business growth.
Two-Tower rankingTri-Tower-XidANN retrievalint8 quantizationCLIP embeddingsCollaborative filteringSwing i2iNearline retrievaleCPM optimizationGMV MaxROASCost Cap biddingNDCG > 0.85recall@k
PythonC++JavaTensorFlowSparkFlinkHiveSQLDockerLinuxFeature pipelinesTraining dataOn-call runbooksPostmortems
01Impact
Numbers that moved.
Cumulative results from the TikTok Shop Ads systems I led or built.
+0%
TTMall ad revenue lift from retrieval improvements
+0%
TTMall ad revenue lift from ranking architecture
0M+
candidates searched by ANN retrieval in under 100 ms
0%+
GMV Max campaign penetration reached through optimization
0
engineers led across ML, product, ads, and commerce delivery
Carousel Ads · launched Q1 2025
$0K+
peak-day revenue for a brand-new ad format with product-aware and GMV Max optimization
02How I work
Turn ambiguity into systems that perform.
Model quality is only part of the craft. The rest is judgment, latency, data, operational discipline, and the confidence to lead through unclear terrain.
-
Find the signal
Separate durable behavior from noisy events before building anything. The best model in the world can't fix the wrong objective.
-
Design for scale
Latency, reliability, and iteration speed are first-class requirements. Ten million candidates in under 100 ms is a design constraint, not a benchmark.
-
Ship measurable impact
Connect model quality to user, advertiser, and marketplace outcomes. If it doesn't move a number someone cares about, it isn't done.
-
Lead through clarity
Roadmaps, runbooks, and postmortems keep teams moving well through unclear terrain. Clarity is a force multiplier.
03Journey
From platform craft to marketplace intelligence.
Keep scrolling — the timeline moves sideways.
-
2014 — 2018
Purdue University
B.S. Computer Engineering · West Lafayette, IN
Computer Engineering foundation with a 3.96/4.0 GPA — where systems thinking and engineering rigor took root.
ThroughlineUnderstand the system, then improve the signal.
Foundation -
2018 — 2020
Salesforce
Software Engineer · Indianapolis, IN
Built reusable web components on AEM for salesforce.com and implemented the analytics framework on Adobe Analytics and Google Analytics 360.
ThroughlineBuild reusable surfaces, then instrument the experience.
Web platform · analytics -
2020 — 2021
Amazon
Software Engineer · Seattle, WA
Designed large-scale retail services and distinctive shopping experiences for customer subsegments, and drove operational excellence across service infrastructure.
ThroughlineRetail systems added the marketplace context behind today's commerce-ads work.
Retail services · operations -
2021 — now
TikTok
Staff Machine Learning Engineer · Tech Lead · Seattle, WA
Leads ML systems for TikTok Shop Ads delivery: retrieval, ranking, Carousel Ads, GMV Max optimization, and engineering quality. Promoted to Staff in January 2025 for technical leadership and cross-team impact.
ThroughlineMarketplace context, ads delivery, ML systems, and technical leadership converged at global scale.
Staff ML · ads delivery · leadership -
Next
What's next?
Open to conversations about ML systems, commerce, ads, and the people who build them.
Say hello
04What I build
Technical depth, translated for humans.
Six systems that define the craft, each with the result it produced.
05Signature work
Proof points with architecture behind them.
Each one is a technical build and a business outcome.
01 / 05
2021 — now
Built TikTok Mall and Shop Ads delivery from 0 to 1, globally.
Led an 8-person team across cross-functional partners to build the delivery-algorithm foundation for TikTok commerce ads, reaching meaningful global scale and revenue impact.
$3M+daily revenue reached
95%+GMV Max penetration
8engineers led
- Signalsuser, product, ad, format
- Delivery0 → 1 global system
- OptimizationGMV Max and bidding
- Outcomemarketplace revenue at scale
- Planned quarterly roadmaps against monetization goals and market fit.
- Collaborated with multiple product, ads, and engineering teams.
- Balanced delivery speed with operational discipline and measurable advertiser value.
02 / 05
Retrieval engine
Expanded retrieval quality across millions of product candidates.
Launched multiple retrieval algorithms: real-time precise retargeting, CLIP-based item-to-item retrieval, item-based collaborative filtering, user-based swing, and organic nearline retrieval.
+55%TTMall ad revenue
+65%advertiser value
<100msANN retrieval latency
- Candidates10M+ product space
- Embeddingstext and image CLIP
- ANN servequantized int8 compute
- Recallefficient high-value set
- Built embedding-based retrieval served through approximate-nearest-neighbor infrastructure.
- Used quantized computation to control latency while preserving retrieval quality.
- Defined retrieval-efficiency metrics such as recall rate@k and percolation rate@k.
03 / 05
Ranking architecture
Designed ranking models that connect relevance, ads value, and product context.
Architected low-latency ranking systems capable of scoring thousands of candidates while optimizing eCPM and marketplace outcomes.
+35%TTMall ad revenue
+40%advertiser value
>0.85NDCG achieved
- Tower Auser and context
- Tower Bproduct and ad
- Tri-Towerformat and Xid signals
- RankeCPM and relevance
- Launched a Two-Tower deep neural LTR model for efficient candidate scoring in under 50 ms.
- Introduced a Multi-Head Tri-Tower-Xid model considering user, product, ad, and format together.
- Optimized ranking for both system latency and advertiser value.
04 / 05
Launched Q1 2025
Launched Carousel Shop Ads with product-aware optimization.
Designed and built the end-to-end delivery flow for Carousel Ads, including image-level selection, ranking, expansion, and AGIC integration.
$900K+peak-day revenue
Q1 2025launch
E2Edelivery flow owned
- Creativeimage candidates
- Selectionproduct-aware ranking
- ExpansionAGIC integration
- DeliveryGMV Max optimization
- Drove image-level selection, ranking, and expansion algorithms for the format.
- Integrated product awareness into delivery optimization.
- Scaled a new format from launch into meaningful peak-day revenue.
05 / 05
Staff promotion · Jan 2025
Raised the engineering bar while leading high-impact ML delivery.
Promoted from Senior to Staff for technical leadership and cross-team impact, with ownership across roadmaps, on-call quality, runbooks, postmortems, and engineering-excellence practices.
Staffpromotion, Jan 2025
Coredelivery on-call owner
X-teamtechnical leadership
- Roadmapquarterly planning
- Qualityrunbooks and on-call
- Reflectionpostmortems
- Cultureshared practices
- Led TikTok Shop Ads core delivery on-call and reliability practices.
- Established engineering best practices that made complex systems more maintainable.
- Turned incidents into stronger shared operating habits through postmortem reflection.
06What people say
In their words.
Sitian is an individual who digs in, learns, and then takes extreme ownership over whatever area you throw at him. He does this while also being a great teammate for his immediate team as well as the broader group. As an example, Sitian went out of his way to run brownbag sessions about various topics for our broader team. He did so with the poise and delivery of a more senior engineer. I wouldn't hesitate to provide a strong endorsement for Sitian.
Sitian was an outstanding addition to our product development team during the summer of 2017. As part of his internship program, Sitian was tasked with building a prototype for a new software application that our organization was planning to launch in the coming months. Not only did he deliver on every detail we asked him to tackle, he impressed everyone with his attention to programming detail, thoughtful product design and a consistent mindset of sharing and seeking feedback. Sitian has a ton of potential to navigate through the engineering ranks as he gains more experience, but frankly he can deliver right of the gates for you on a number of fronts. I would highly recommend him to any employer that can offer an environment of sharing and collaboration, fast-paced product development and/or an opportunity where you need a well-rounded, humble, and talented engineer with potential to grow.
07Off the clock
Outside the system.
The non-work side — worth a click or two.
Chinese 8-ball
Heyball regular.
Part of the Seattle Heyball club scene. Give me a table and a rack, and the ranking model can wait an hour. Click the table to break.
Dota 2, first and always.
CS:GO and Hearthstone for when the queue is long.
- Dota 2
- CS:GO
- Hearthstone
Long arcs, real stakes.
Naruto, Jujutsu Kaisen, and Pokémon — stories that stick around long enough to grow with you.
- Naruto
- Jujutsu Kaisen
- Pokémon
Lives with a rabbit.
Quiet roommate, strong opinions, zero interest in ads ranking.
Perpetually poking at new things.
New models, new tools, new gadgets. If it's clever, I want to take it apart.
--:--
Greater Seattle Area. Coffee, rain, and a lot of ranking experiments.