Data Science Lead · Recommendation & Advertising · Experimentation & Causal Inference · GenAI Agents
I work at the intersection of business and algorithms: turning ambiguous business questions into testable data questions, then shipping the models, experiments and systems that answer them.
Each area comes with the code, issues and talks behind it.
| Area | In practice | Code, issues & talks |
|---|---|---|
| GenAI & agents | ReAct + tool-calling agents, tool schema validation and fallback, answer-evaluation pipelines, end-to-end observability (token / latency / cost / hit rate), proactive outreach systems | OpenAI-Data-Analytics ★30 — data-agent workflows AutoKaggle — agent-driven ML pipelines transformers #39077 GeneralDataAgent - data agent, building |
| Business insight & growth | User profiling and segmentation, retention / conversion analysis, content and user value scoring, indicator systems, opportunity discovery from billions of behavioral events | 《AI + 内容精细化运营下的产品增长》 · 人人都是产品经理大会 2023 |
| Experimentation & causal inference | End-to-end experiment design (assignment → metrics → guardrails → diagnosis → attribution), two-sided-market A/B tests, propensity score matching / IPW, uplift modeling, ablation-based validation | 《数据科学在音乐推荐中的实践和应用》 · msup A2M 2022 |
| Ads & campaign optimization | Built an advertiser-facing (label-side) promotion platform from 0 to 100M+ RMB ARR; ranking evolution XGBoost → DIN / DCN / STAR; budget allocation by incremental lift | 《内容评估理解和流量精细化运营》 · DataFunTalk 2021 |
| Recommendation & multimodal | Retrieval / ranking / re-ranking, EPNet-family fine ranking, CLIP + Qwen3-Embedding representations, Faiss retrieval, cold start, diversity constraints | VoicePrintReco ★80 — speaker recognition on ALIZE MFCCandDTW ★41 — keyword spotting htk_3_4_1 ★47 - ASR |
| Systems & tooling | Protocol servers, LaTeX tooling — the C++ years that shaped how I read systems today | xsmtp ★45 · xpop3 ★4 · HttpServer ★1 zjuthesis ★11 — LaTeX thesis template |
- 0 → 100M+ RMB ARR: led the B-side music promotion platform serving major labels and thousands of independent artists; twice awarded the company-level Business Breakthrough Award.
- +10% inventory utilization and campaign completion rate, with no material regression on listener experience — via uplift-based budget allocation and cross-domain ranking.
- +5% impression→play conversion, listening time and saves on long-audio fine ranking.
- +3% play penetration from multimodal (CLIP + Qwen3-Embedding + Faiss) retrieval.
- +20% answer accuracy turning a workflow-style music assistant into an autonomous tool-calling agent.
- Discovered and boosted breakout hits (e.g. 《一路生花》) with a hit-potential scoring model built on lyrics / audio signals and leading-indicator user behavior.
Open to conversations about data science leadership, experimentation infrastructure, recommendation / advertising algorithms, and agent systems.
📫 ibillxia AT gmail DOT com · 🌐 ibillxia.github.io · 📍 Shenzhen, China
中文简介
高级数据科学家 / 数据科学负责人。
长期做三件事：从海量用户行为里挖业务洞察，用严谨的实验和因果推断衡量效果，再把结论落成模型和系统。
主导 B 端投放平台从 0 到亿级 ARR；用双边市场 A/B 与 Uplift 建模做增量预算分配，库存利用率与任务完成率 +10%；迭代召回与精排模型，播放转化与时长 +5%；把音乐助手升级为 ReAct + Tool Calling 的自主 Agent，准确率 +20%。
技术栈：Python / SQL（HiveSQL）、scikit-learn、statsmodels、PyTorch、Spark / Flink / Kafka、Faiss、Go。