Machine Learning
Supervised, ranking, and graph models from baseline to production — with rigorous evaluation and SHAP-level explainability.
Dr. Priya Nair · Open to new roles
I turn messy data into models that move the metric.
I'm Priya — a data scientist and ML engineer with nine years shipping models to production. From recommender systems to NLP pipelines, I build the boring-reliable kind of machine learning that survives contact with real users.
120+Models shipped to production
$8MIncremental revenue driven
9yYears in applied ML
Tools and frameworks I work with every day
Python
PyTorch
scikit-learn
Spark
dbt
Airflow
AWS
Selected work
See all projects
Recent projects
01
+18% recall
Churn Sentinel
ML · Retention
A gradient-boosted churn model that flags at-risk accounts two weeks earlier, lifting recall 18 points over the prior baseline.
- XGBoost
- Feature Store
- SHAP
02
-63% triage time
TicketRouter NLP
NLP · Support
A transformer classifier that auto-routes support tickets, cutting human triage time by nearly two-thirds.
- Hugging Face
- BERT
- FastAPI
03
+24% conversion
Atlas Recommender
ML · Personalization
A two-tower recommender serving 40M daily impressions, lifting click-through conversion 24% in a controlled A/B test.
- PyTorch
- Faiss
- Redis
04
0.94 AUC
FraudGraph
ML · Risk
A graph-based fraud detector scoring transactions in real time at 0.94 AUC, saving an estimated $2.1M annually.
- GNN
- Kafka
- Neo4j
05
99.2% accuracy
DefectLens Vision
Vision · Manufacturing
A CNN defect detector for a production line, reaching 99.2% accuracy on edge hardware with sub-50ms inference.
- CNN
- ONNX
- OpenCV
06
+31% forecast acc.
Demand Almanac
Analytics · Forecasting
A hierarchical demand-forecasting system that improved accuracy 31% and cut excess inventory across 1,200 SKUs.
- Prophet
- LightGBM
- dbt
About
Rigorous models, honest about uncertainty.
Based in Toronto and working remotely, I help product and risk teams ship machine learning that holds up in production. I care as much about clean evaluation and monitoring as I do about the model itself — a 0.99 offline AUC means nothing if it quietly rots after launch.
- Problem-framing before model-building
- Reproducible pipelines, versioned data
- Honest metrics, calibrated confidence
Expertise
What I do best
A focused toolkit, sharpened across fintech, e-commerce, and industrial ML.
NLP
Transformers, classification, and retrieval pipelines that read, route, and summarize text at scale.
Analytics & Forecasting
Causal inference, experimentation, and time-series forecasting that turn dashboards into decisions.
Computer Vision
Detection and classification models optimized for edge inference and real production constraints.
MLOps
Feature stores, CI/CD for models, drift monitoring, and the plumbing that keeps ML alive after launch.
Experimentation
A/B test design and analysis that separates real lift from noise, so the metric you move is the metric that matters.
By the numbers
Nine years, measured in shipped models
120+Models in production
40MDaily predictions served
9yYears in applied ML
3Patents & papers
Kind words
What teams say
Priya is the rare data scientist who ships. She framed the problem, built the model, and stayed through deployment until the metric actually moved.
Marcus LeeVP Engineering, Atlas
Our fraud losses dropped within a quarter. What impressed me most was how carefully she communicated the model's limits, not just its wins.
Elena RossiHead of Risk, Verde Pay
She translates between research and product fluently. The NLP pipeline she built is still humming two years later with almost no maintenance.
David OkaforCTO, Pulse Support