I build production ML systems that deliver customer impact at scale by any means necessary.
- Sometimes this means getting in the weeds wrangling raw data; training and evaluating models; or setting up super reliable high-throughput, low-latency, highly-observable online/offline ML feature and model serving systems1.
- Other times this means sitting in meetings with business, product, and technical stakeholders from all corners of the company, as well as partners and colleagues across the industry, to define the right problems to tackle and the right things to build.
- No matter what, it always means listening to the people we're serving, making sure we're solving the most important problems, and thinking beyond what's possible to what is actually needed to add the most value to our customers2.
I'm currently focus-firing all of the above as a Staff Machine Learning Engineer at Mercari, Japan’s largest C2C e-commerce marketplace, to support the world-class AI teams and applications across the company.
- FOSSASIA Summit 2023: The Bumpy Road of Bringing a Machine Learning Model From Development to Production - Part 2: Infrastructure, Testing, Productionalization, and Monitoring3
- Berlin Buzzwords 2023: Building MLOps Infrastructure at Japan's Largest C2C E-Commerce Platform3
- FOSSASIA Summit 2024: Unleashing the Potential of AI Through Search Ranking: How Mercari Uses Data & Science to Drive Marketplace Growth4
- Weights & Biases Fully Connected 2024 (San Francisco): How Mercari is Using Gen AI to Define the Future of Japanese C2C E-commerce
- Berlin Buzzwords 2024: Robust AI Search Ranking for Radical C2C Marketplace Growth4
- Weights & Biases Fully Connected 2024 (Tokyo): The Gen AI Powering the Next Generation of the Mercari Marketplace
- [Coming Soon] FOSSASIA Summit 2025: LLMOps for Eval-Driven Development at Scale5
Footnotes
-
My current personal record was for an AI search ranking system at Mercari:
- Offline data & model training pipelines that processed petabytes of data.
- Online serving system that handled nearly 6K RPS of search traffic with average e2e latency <35ms (p95
<100ms) and five 9's uptime.
- Streaming real-time feature ingestion pipeline that processed over 100K client events per second and over 2.3 terabytes of data per day; significant optimization for a monthly operational cost of $567/mo.
- Feature store that served over 3.5M read/write RPS with average read latency <3ms (p95 <10ms).
- Model server that handled nearly 6K RPS with average read latency <13ms (p95 <38ms).
-
If A/B test results are at least a decent proxy for customer value, the work I spearheaded at Mercari in AI search ranking led to a projected $50M/year revenue increase. ↩
-
w/
@chingisooinar↩ ↩2 ↩3 -
w/
@jehandadk↩