Zafarullah Mahmood
Designing human-centred AI systems with and for care.
About Me
I am an NLP researcher and engineer who enjoys building systems that feel thoughtful to the people who rely on them. I prioritize rigorous experimentation and delightful product experiences.
Beyond work, I am curious about philosophy of science, linguistics, and behavioural economics.
Timeline
ML Engineer (Contract)
Pinterest / Innodata Inc.
Currently working on improving hybrid search relevance for Women's Fashion business vertical. Responsibilities include context engineering, automated prompt fine tuning, and automating data labeling workflows using LLM-as-a-Judge.
Master of Applied Science in Computer Engineering
University of Toronto
Thesis: "A Fully Generative Counsellor Chatbot for Smoking Cessation and LLM-Based Synthetic Smokers"
Natural Language Processing Engineer
Dialpad Canada Inc.
I worked closely with product and research teams to deploy speech and language models that powered millions of customer conversations. My contributions include shipping a bilingual sentiment classifier that matched Spanish performance without sacrificing English accuracy, and developing an n-gram boosting approach that improved domain keyword recognition by 26%.
I also unified punctuation and casing into a single BERT model, cutting inference latency by 4% and reducing errors.
Data Scientist
Exzeo Software, India
I designed data products that helped insurance teams respond quickly to customer needs, including an AWS-hosted pipeline that transcribed voice calls and surfaced insights for claim assessors.
Bachelor of Technology in Computer Engineering
Jamia Millia Islamia, India
Skills at a Glance
Programming Languages
Deep Learning & NLP
Generative AI Systems
MLOps & Cloud
Product Engineering
Publications
- AutoAnnoMI: A Framework to Automate the Annotations of MI Conversations using LLMs (2024)
- A Fully Generative Motivational Interviewing Counsellor Chatbot for Moving Smokers Towards the Decision to Quit (Findings of the Association for Computational Linguistics: ACL 2025)
- N-gram Boosting: Improving Contextual Biasing with Normalized N-gram Targets (arXiv preprint, 2023)
- Avengers, Ensemble! Benefits of ensembling in grapheme-to-phoneme prediction (SIGMORPHON, 2021)
Projects
- Exploring Proposer-Acceptor Disparity in Dynamic Matching (2023): Analyzed utility disparity in the popular problem of dynamic stable matching by simulating deferred acceptance with agent arrivals; proposed a time-annealed cutoff utility that narrows the gap without reducing match rates.
- TensorFlow Speech Recognition Challenge (2018): Participated in a Kaggle challenge to build a CNN to classify 30 short speech commands.
- Benchmarking Batch Renormalization (2017): Implemented BatchReNorm1d module in PyTorch and benchmarked its performance.
- ShapSum: Shapley Value-Based Multi-Document Summarization (2016): Developed a novel multi-document summarization method.
- Random Acts of Pizza (2016): Predicted whether a Reddit user would receive a pizza based on their request.
- TidyWord (2016): Text cleaning software in C for removing less frequent words from large documents.
Certifications
- Develop Generative AI Applications (IBM, 2025)
- Creating Behavioral Change (Wesleyan University, 2023)
- Introduction to Psychology (University of Toronto, 2023)
- Philosophy of Science (University of Pennsylvania, 2023)
- Miracles of Human Language: An Introduction to Linguistics (Leiden University, 2021)
- Introduction to Deep Learning (HSE University, 2018)
- Mathematics for Machine Learning Specialization (Imperial College London, 2018)
- Algorithms Specialization (Stanford University, 2017)
- Machine Learning (Stanford University, 2016)
- Introduction to Programming Using Python (MIT, 2015)
Awards
- Edward S. Rogers Sr. Graduate Scholarship, University of Toronto (2024)
- Best Kernel Award, Kaggle (2018)
- Deloitte Innovation Award, Indian Space Research Organisation (2017)