Applied ML
Search, recommendation, ranking, experimentation, and product measurement.
Myeonghwa Lee
Product-minded AI/ML Engineer
I build data and AI products end to end—from instrumentation and modeling to production delivery and measurable iteration.
Across medical-aesthetics discovery and mobile-game analytics, I have shipped recommendation and ranking systems, production data and ML platforms, and agentic AI products.
Verified evidence
Outcomes with their context intact.
- +7%Purchase conversionWishlist recommendation
- +23%Click-through rateInterest-category product experience
- +6.59%Purchase conversionInterest-category product experience
- PLOS ONEPublished researchFirst-listed, equal-contributing author
End-to-end range
Three disciplines, one product loop.
The work spans model behavior, the systems around it, and the AI products people ultimately use.
Data & ML Systems
Event design, Databricks, dbt, Airflow, production ML operation, and reliability.
AI Products
Multimodal RAG, agent workflows, evaluation, orchestration, and decision intelligence.
Selected work
Evidence-led case studies.
Each project traces the problem, system, delivery path, and available outcome without separating the model from the work around it.
Healing Paper
Gangnam Unni — Search, Recommendation & Ranking
Built the data, models, delivery paths, and measurement loops behind a medical-aesthetics discovery experience.
Challenge
Connected fragmented discovery signals to relevant recommendations and a measurable search experience.
My role
End-to-end ownership across event definition, search data marts, modeling, batch/real-time delivery, deployment automation, monitoring, and product iteration.
Approach
- Defined search events and sessions, then transformed raw logs with EMR/PySpark into dbt/Airflow marts.
- Built and shipped recommendation and ranking models through batch and real-time paths.
- Created dashboards, daily reporting, and anomaly detection to make the product loop observable.
- Iterated with product teams against click, engagement, and purchase metrics.
Outcomes
- Wishlist recommendation purchase conversion +7%.
- Community recommendation purchase conversion +2% and likes +14%.
- Interest-category experience click-through rate +23% and purchase conversion +6.59%.
- Ranking page achieved 40% click-through rate.
Capabilities
- Recommendation & ranking
- Search data products
- Batch & real-time delivery
- Experimentation & measurement
Bagelcode
NebulaScope — Tableau Embedded Centralized Analytics & AI Platform
Built the platform layers that turn analytics events into a Tableau Embedded product—and then into multilingual AI decision intelligence.
Challenge
Created a shared path from consistent product instrumentation to trusted analysis across a multi-game business.
My role
Ownership across Analytics Event design, Databricks mapper/metadata work, all NebulaScope dbt models and pipelines, Airflow/ETL operation, all NebulaScope Tableau workbook work, and NebulaScope Insight.
System architecture
From event design to interpretation
- Instrument
Shared Analytics Event design and governance.
- Standardize
Databricks mapper and metadata centralization.
- Model
dbt models, dependencies, and Airflow ETL.
- Deliver
Tableau Embedded analytics across acquisition, retention, monetization, progression, engagement, and operations.
- Interpret
NebulaScope Insight AI decision intelligence.
Reliability & context control
NebulaScope Insight: reliable AI analysis, not a summary button.
Designed an AI analysis pipeline that combines app-specific dashboard interpretation, cross-dashboard prioritization, and multilingual delivery while preserving useful output through partial failures.
Reliability by design
- Separate dashboard and overall passes.
- Durable priority output before translation.
- Bounded concurrency and isolated failures.
- Failure-rate gates, rerun/skip behavior, and forced regeneration.
- Streaming analysis-tool loops with retries, backoff, truncation handling, token accounting, and robust response extraction.
Scale & context control
- Korean, English, and Vietnamese output.
- Compact TSV representations.
- Coverage-aware dimension trimming.
- App-direction injection for relevant context.
Outcomes
Established a shared analytics path from governed events through embedded delivery to multilingual AI interpretation.
Capabilities
- Analytics Event governance
- Databricks
- dbt & Airflow
- Tableau Embedded
- AI decision intelligence
Bagelcode
AEGenie — Analytics Event Lifecycle Platform
Turned a hackathon idea into an internal AI product for designing and validating analytics events.
Challenge
Reduced the manual work and inconsistency involved in translating product specifications into production-ready event designs.
My role
Built the project end to end.
Approach
- Multimodal ingestion of product specifications.
- Retrieval over prior Analytics Event documents and code.
- Vector search, RAG, and a multi-step agent workflow.
- Natural-language event generation and validation.
- API services, containerization, CI/deployment work, and security hardening.
Outcomes
Advanced from an internal hackathon prototype into a real internal project.
Capabilities
- Multimodal ingestion
- Retrieval & RAG
- Agent workflows
- Event validation
- API delivery
Experience
Career evidence in context.
Product ML, analytics platforms, and AI systems developed through continuous ownership from June 2022 to the present.
Bagelcode
Data Scientist / AI & ML Engineer
January 2025–Present
Projects NebulaScope · AEGenie · Voyager
- Built NebulaScope's centralized analytics path end to end: Analytics Event design, Databricks mapper/metadata centralization, dbt models and Airflow pipelines, and the Tableau Embedded delivery layer.
- Designed NebulaScope Insight, a multilingual AI decision pipeline with separate dashboard and overall passes, priority-aware output, bounded concurrency, partial-failure isolation, retry controls, and context optimization.
- Built AEGenie end to end, combining multimodal product-spec ingestion, retrieval, RAG, agent workflows, validation, APIs, containerization, and delivery hardening; the hackathon prototype progressed into a real internal project.
- Operated and improved Voyager's production SKAdNetwork ML system across hourly data/inference, daily attribution, and weekly training flows, strengthening repeatability, retry/logging behavior, Delta merge reliability, concurrent-delete handling, and month-boundary correctness.
- Operated daily Airflow workflows and recurring Databricks ETL across production analytics systems.
Healing Paper
Data Scientist, Gangnam Unni app
June 2022–January 2025
Product Gangnam Unni
- Built and iterated recommendation and ranking products across batch and real-time delivery, improving wishlist recommendation purchase conversion by 7% and community recommendation purchase conversion by 2% while increasing likes by 14%.
- Developed an interest-category inference experience that increased product click-through rate by 23% and purchase conversion by 6.59%.
- Built the search data product end to end—from event and session definitions through EMR/PySpark transformation, dbt/Airflow marts, dashboards, daily reporting, and anomaly detection.
- Automated model delivery and measurement workflows, enabling repeated product experiments and KPI-driven iteration.
- Built a ranking experience that achieved a 40% click-through rate.
Research & education
Research grounded in social data.
Academic work in computational social science continues to shape how I frame evidence, measurement, and model behavior.
Publication
Do political parties always prefer loyalists? Evidence from South Korea
First-listed and equal-contributing author; data curation, resources, visualization, and original-draft writing.
Education
M.S. · Graduate School of Culture Technology, KAIST
Social data science, data mining, graph mining, NLP, GNNs, and stochastic processes.
Capabilities
Tools grouped by the work they enable.
A practical stack spanning applied modeling, AI workflows, data and ML systems, and production delivery.
- Applied ML
- Recommendation · Ranking · Search · Forecasting · Experimentation
- AI Systems
- RAG · Multimodal Workflows · Agents · Evaluation & Validation
- Data & ML Systems
- Python · SQL · Spark · Databricks · dbt · Airflow · MLflow
- Production
- AWS · Docker · APIs · GitHub Actions · Monitoring
About & contact
Useful systems keep learning after launch.
I like problems where model quality is only the beginning. My best work connects instrumentation, data systems, production delivery, and product feedback so that a useful idea can keep improving after launch.