Experience
Amazon - Applied Scientist (2022–Present), Vancouver, Canada
- Ideated and built an internal LLM experimentation and evaluation toolkit, demonstrated faster time-to-production, then partnered with engineering teams to scale it across multiple production generative AI applications
- Built LLM judges for automated evaluation:
- Reference-based fact-checking judges, designed to be use-case-agnostic and reused across projects
- Judges that estimate continuous values via prompting, applied to a new domain
- Led cross-team science work on an LLM-based recommender in partnership with engineering
- Mentored applied scientists and interns on experiment design and coding
Elpha Secure - Lead Applied Scientist (2021–2022), Vancouver, Canada
- Led cross-team efforts building ML products for cybersecurity, including classifiers for encrypted command detection, suspicious login detection, and URL phishing
- Designed an ML pipeline to speed up experimentation and deployment, accelerating how new ML products were built and shipped
- Automated model deployment with lightweight, scalable Kubernetes scripts
Olyns - Machine Learning Consultant (2021), San Francisco, USA
- Verified existing ML models and supported their deployment on AWS
Georgian - Applied Data Scientist (2020–2021), Toronto, Canada
- Performed deep error analysis on an existing ML product to help business and engineering stakeholders understand its behavior
- Designed NLP models for hierarchical text classification in the financial and marketing domain
EhsAI - NLP and Machine Learning Scientist (2019–2020), Vancouver, Canada
- Built NLP deep learning models for document intelligence in the environment, health, and safety domain
- Designed data augmentation and synthetic data generation methods to reduce manual labeling
- Built an ML framework for deploying and evaluating pretrained language models for classification tasks
Education
PhD, University of British Columbia (2012–2019), Vancouver, Canada
- Designed recommender systems for data exploration, published at ICDE and UMAP
- Built deep learning models (LSTM/CNN) to predict SQL query properties before execution, published at SIGMOD
- Developed statistical methods for estimating SQL query answer properties in data integration settings, published at EDBT
- Facilitating User Interaction With Data, PhD Thesis · slides
MSc, Dalhousie University (2009–2012), Halifax, Canada
- Built a statistical topic model for question answering and automatic tagging in community Q&A archives
- Studied expertise modeling to route new questions to the right experts
- Probabilistic Modeling in Community-Based Question Answering Services, MSc Thesis
BSc in Computer Software Engineering, Isfahan University (2004–2008), Isfahan, Iran
- Built autonomous soccer-playing agents for the “UI-AI” RoboCup team over three years - my entry point into AI
Publications
Workload-Aware Query Recommendation Using Deep Learning (EDBT, 2023)
paper
This paper studies how workload context can improve recommendations for database queries. It uses deep learning to suggest useful queries while accounting for the patterns and needs represented in a user's existing workload.
Facilitating SQL Query Composition and Analysis (SIGMOD, 2020)
paper · dataset
This work presents techniques for helping people construct and understand SQL queries. It focuses on supporting the full interaction around a query, from composing a statement to examining its behavior and results.
Improvement of SQL Recommendation on Scientific Database (SSDBM, 2019)
paper
This paper investigates how SQL recommendations can be improved for scientific databases. The work considers the specialized workloads and query patterns found in scientific data analysis to make recommendations more useful to researchers.
A Generic Top-N Recommendation Framework for Balancing Accuracy, Novelty, and Coverage (ICDE, 2018)
paper · slides
This paper introduces a general framework for designing Top-N recommenders that balance several competing goals. In addition to accuracy, it considers whether recommendations are novel and whether they provide broad coverage of the available items.
Bridging the Gap Between User-Centric and Offline Evaluation of Recommendation Systems (UMAP, 2018)
paper
This work examines the relationship between offline recommender-system metrics and the experience of real users. It highlights why evaluation should connect algorithmic measurements with user-centered outcomes rather than relying on a single offline score.
Facilitating User Interaction With Data (PhD@VLDB, 2017)
paper
This research explores ways to make data systems more approachable for people who need to ask questions of data. It brings together query composition, recommendation, and answer analysis to support users throughout an interactive data exploration process.
Extracting Aggregate Answer Statistics for Integration (EDBT, 2015)
paper · slides
This paper studies how aggregate statistics about answers can be extracted and used to support data integration. The approach helps systems reason about query answers and combine information in a way that is useful for subsequent analysis.
Finding Expert Users in Community Question Answering (WWW, 2012)
paper · dataset
This work investigates how to identify expert contributors in community question-answering archives. It uses signals from user activity and answered questions to help distinguish knowledgeable participants who can provide reliable guidance.
Modeling Community Question Answering Archives (NeurIPS, 2011)
paper · poster · dataset
This paper models the content and activity found in community question-answering archives. The goal is to better understand how questions, answers, and contributors relate to one another in these collaborative knowledge spaces.
Datasets
StackOverflow Q&A Archive Dataset
download
A dataset of Stack Overflow questions, answers, and duplicate-question pairs. Tags were deliberately selected by frequency and co-occurrence to balance easy and hard cases, and the test set uses Stack Overflow's own community-identified duplicate questions as a gold standard for answer retrieval, rather than requiring a user study. Built for the "Modeling Community Question Answering Archives" and "Finding Expert Users in Community Question Answering" papers and used in my MSc thesis.
SDSS Query Workload Dataset
dataset: TBD
A SQL workload dataset extracted from Sloan Digital Sky Survey (SDSS) query logs: sampled one query per session and deduplicated down to 618,053 unique query statements from an original 194 million log entries across roughly 1.6 million sessions. Built for the "Facilitating SQL Query Composition and Analysis" paper.
Posters and Demos
Facilitating Data User Interaction With Data (Microsoft Research AI Breakthroughs Workshop, 2019)
poster
This demonstration presents an interactive approach to helping people work with data. It emphasizes practical user interaction and shows how research ideas can support more direct and productive data exploration.
Facilitating Data Exploration, Query Composition, and Query Answer Analysis (NWDS, 2018)
poster
This demo brings together several stages of a data workflow in one user-facing experience. It supports exploring data, composing queries, and examining answers so that users can move more easily from an initial question to an informed result.
Personalized Top-N Recommendation for Promoting Long-Items (WIML/NeurIPS, 2017)
This work considers how personalized Top-N recommendation can help surface longer or less frequently selected items. It focuses on recommendation strategies that account for user interests while improving visibility beyond the most obvious short-list choices.
Talks
- Generative AI Tutorial - Amazon Machine Learning Conference (AMLC)
- Facilitating SQL Query Composition and Analysis - SIGMOD Conference, Portland, OR (2020)
- Facilitating User Interaction With Data - Huawei Noah's Ark Lab, Toronto (2019)
- Facilitating User Interaction With Data - Thomson Reuters, Toronto (2019)
- Bridging the Gap Between User-Centric and Offline Evaluation of Recommendation Systems - UMAP, Singapore (2018)
- A Generic Top-N Recommendation Framework for Balancing Accuracy, Novelty, and Coverage - ICDE, Paris (2018)
- Facilitating User Interaction With Data - PhD@VLDB, Munich (2017)
- Extracting Aggregate Answer Statistics for Integration - EDBT, Brussels (2015)
© 2026 Zeinab Zolaktaf
