Mustapha Unubi Momoh | Machine Learning Engineer

Hi, I'm Mustapha Unubi Momoh.

a/an

I solve real-world problems with machine learning

About

Machine Learning Engineer

I build production recommender systems, ML infrastructure, and applied GenAI systems on AWS, Kubernetes, and NVIDIA tooling.

My recent work focuses on ranking systems, model serving, feature retrieval, offline evaluation, retraining workflows, and operational reliability for machine learning systems.

Core Focus

Production recommender systems
ML serving and infrastructure
Offline evaluation and monitoring
Generative AI applications

Stack

Languages: Python, R, Bash
ML: TensorFlow, PyTorch, NVIDIA Merlin
Data: NumPy, Pandas, OpenCV, PyMC
Cloud: AWS, GCP, Azure, Docker, Kubernetes

Open to challenging opportunities in recommender systems, production ML, GenAI, and MLOps.

Featured Projects

Production ML Case Study

Production Multistage Multimodal Recommender on Amazon Elastic Kubernetes Service (EKS)

End-to-end recommender system with candidate generation, ranking, reranking, filtering, feature caching, and Triton serving on Kubernetes.

System

Two-Tower candidate generation, DLRM ranking, diversity reranking, and seen-item filtering.
Cold-start handling with feature masking, context-aware recommendations, CLIP embeddings, and Sentence-BERT embeddings.
Serving stack with Amazon EKS, NVIDIA Triton, Feast, FAISS, Kubeflow, and Valkey-backed Bloom filters.

Impact

99.7%: item feature lookup latency reduction
54%: end-to-end latency reduction
310%: throughput improvement

Interactive Demo

The demo exposes the serving system as a recommender UI with controls for user ID, device type, time of day, and Top-K results, making the personalization inputs and scored outputs visible without opening the code.

Interactive multistage recommender system demo with user, context, Top-K controls, scores, and recommendation cards

Triton Serving Ensemble

A single client request flows through a Triton ensemble: context preprocessing, Feast-backed user lookup, NVTabular transforms, Two-Tower retrieval with FAISS, Bloom-filter seen-item removal, item feature lookup, DLRM ranking, and final softmax sampling.

Triton serving graph for multistage recommender with context, retrieval, filtering, item features, ranking, and response stages

Feature Caching Optimization

Profiling showed per-request item feature lookup caused hundreds of Redis/Valkey round trips through Feast. Loading item features into an in-memory NumPy cache at model initialization reduced lookup latency from 195 ms to 0.5 ms.

Training, Deployment, and Monitoring

The MLOps flow keeps training and serving on Amazon EKS: Kubeflow prepares data and models, artifacts are persisted to Amazon EFS, NVIDIA Triton Inference Server serves the 14-model ensemble, and Prometheus/Grafana track utilization, throughput, and latency for capacity planning.

MLOps architecture for multistage recommender on Amazon EKS with Kubeflow, Triton, Prometheus, Grafana, GPU nodes, and CPU nodes

Tools: Amazon EKS, NVIDIA Merlin, NVIDIA Triton, Feast, FAISS, Kubeflow, Redis/Valkey, CLIP, Sentence-BERT

TDS Article Medium Article Demo Code

ML Infrastructure Case Study

Recommender System with Continuous Retraining on Amazon Elastic Kubernetes Service (EKS)

DCN-based ads-ranking recommender deployed on Amazon EKS with monitoring, drift-triggered retraining, and autoscaling.

System

Built and deployed a DCN-based ads-ranking recommender system on Amazon EKS.
Triggered continuous retraining when model performance, based on AUC-ROC, drifts below a defined threshold.
Designed the project around serving, monitoring, retraining triggers, and server autoscaling.

Focus

Production ranking architecture
Model performance monitoring
Kubernetes-native deployment

Tools: Amazon EKS, NVIDIA Merlin, HugeCTR, Kubernetes, model monitoring, autoscaling

Article Code

Experience

Pixite Inc.

Machine Learning Engineer (Recommender systems)

Designed and proposed recommender-system architecture options on AWS and GCP, evaluating tradeoffs in training speed, inference latency, delivery timelines, and operating costs across data ingestion, model training, and inference.
Collaborated with the product team to define data and ranking requirements for personalized search and recommendation features for Pigment app.
Collaborated with engineering to train recommendation models for Pigment app enabling homepage content personalization for millions of users.
Led discussions around recommendation request/response caching to optimize performance, including evaluating trade-offs between different cache types.
Tools: Recommendation algorithms, Vertex AI, and Cloud functions

November 2024 – July 2025 | United States, Remote (Contract)

AWS

AWS Community Builder in Machine Learning and GenAI

Fostering a learning community around AWS services and Machine Learning.
Tools: AWS services, Machine Learning, Generative AI

February 2024 - present | Waterloo, Canada

EveryRate

Data Scientist (OCR, ETL, and Automation)

Designed and deployed an ETL pipeline to extract mortgage rates from structured documents using Azure AI Document Intelligence, Azure Functions, and Blob triggers.
Benchmarked OCR pipeline tools, including Amazon Textract, Google Document AI, Azure AI Document Intelligence, and vision-language models for tabular data extraction.
Automated document processing with blob-triggered functions and upserted extracted mortgage-rate data into PostgreSQL for application use.
Tools: Azure AI Document Intelligence, Azure Functions, Blob Storage, PostgreSQL, Amazon Textract, Google Document AI

May 2024 – December 2024 | Vancouver, Remote

Stealth Startup

Machine Learning Engineer (DS Team lead)

Trained, packaged, deployed deeplearning models for spoofing verification for credit card and spend management companies.
Worked with the VP of Engineering to set up API gateways, and also collaborated on writing the API specification and technical report that detail the benchmarking results.
Led the Data Science team in pitching to two corporate credit card and spend management companies with positive feedback.
Tools: Python, AWS Lambda, Sagemaker Endpoint, tensorflow serving, SQS, Docker, API Gateway

March 2024 – May 2024 | United States, Remote (Contract)

contracts

Generative AI Engineer (expert-vetted)

Worked as a Generative AI consultant for a Copilot development for a Visual programming language.
Built AWS well-architected solutions for startups and companies for usecases including Medical GenAI, Injury Claims LLM-assisted processing, Image upscaling, and OCR for check management.
Worked on a POC of an AI shopping Assistant similar to Shopify’s shop.app but tailored to the client’s inventory.
Worked on Beauty Retail Generative AI POC using PaLM-2, Stable Diffusion, and Vertex AI.
Worked on Causal understanding of REM sleep, Deep sleep, and Sleep latency project with TabNet, SHAP, and PyMC
Tools: AWS Bedrock, Large Language Models (Titan), text embedding, Vector DBs, Amazon Kendra, Streamlit, AWS EC2, Stable Diffusion, GCP, vertex AI, AI agents, Knowledge graphs, Amazon Neptune, neo4j, Amazon kendra, Entity extraction, Intent recognition, Explainable AI with SHAP, Bayesian Causal Inference, and Machine Learning

April 2023 - present | United States (NYC) | Canada (Remote)

Karaam Analytics

Founder

Liased with clients to identify their needs and goals for Data Analytics.
Trained individuals and corporates on SQL and Python for Data Science.
Tools: Python, Data science

June 2020 - present | Lagos, Nigeria

Selected Projects and Hackathons

                
2024 NVIDIA AI Hackathon: AI Assisted Documentation Review Review and Update
                  AI Assisted Documentation Review Review and Update Application using AWQ Quantized 13B llama and TensorRT-LLM
                
AccomplishmentsTools: llama-2, Nvidia TensorRT, TensorRT-LLM, Quantization, Streamlit, docker, Nvidia RTX 4090
Launch app and Login with your Atlassian Confluence Credentials.
Your documentation/articles in Confluence space will be auto downloaded and indexed
Chat with the documentation or
Create new content by providing a title, edit the generated content, and publish

Retrieval Augmented Generation with AWS Bedrock, Kendra, and Amazon Titan 
                  Retrieval Augmented Generation with AWS Bedrock, Kendra, and Amazon Titan for content and slides generation
                
AccomplishmentsTools: AWS Bedrock, AWS Kendra, EC2, Amazon Titan model, Prompt Engineering, Amazon s3
Clone the repo
Launch the application and create long form articles or short powerpoint slides

GenAI app for a Beauty Retail
                  A GenAI Proof of Concept for a Beauty Retail
                
AccomplishmentsTools:Python, Vertex AI, PaLM-2, Finetuning Stable difussion, Large Language Model (chat-bison) finetuning, model deployment
Automate Content creation for a beauty retail website
User purchase histories and interests, along with the query are used to generate branded product and model images in various scenes, before and after transformation photos, alongside relevant descriptions on the home page
User provides basic information about them such as age, race, hair type and length and a prompt is constructed automatically to generate suitable hair product, beauty model, and transformation photos

Medical Decision Support
                  Machine Learning Clinical Decision Support System Proof of Concept with LIME and Decision Trees.
                
Accomplishmentsengineered features such as speech speed, average characters, average nouns, sentiments from intervioew recording and transcripts
trained a decision tree classifier for detecting the likelihood of depression
used Local Interpretable Model Agnostic Explanations (LIME) to produce local feature contributions and Visualizations for interpretable ML
app can generate and display prediction probabilities, decision trees, LIME plots, and Feature importance on the interface
users can generate a short medical report with their assessment

Interactive Text Label Explorer
                  An Interactive Dashboard for Text Label Exploration.
                
Accomplishments
                  The preprocessing steps include:
                  creating word embeddings.
Projecting the embeddings vector to 2D plane using dimensionality reduction techniques (17 of them used in the project)
Topic modeling to produce clusters based on topics.

                  The dashboard allows users to interactively explore the data and labels in different panels including:
                  label-based groupings view
topic-based groupings view
top sentences view
top words and word cloud view

                  Based on findings from the explorations, the user can select data for review directly from the scatterplots. The selected data can be downloaded by clicking on a button. Please watch the demo video for more details.
                
Microsoft Responsible AI Hackathon - Deeplearning Assisted Diagnosis of Primary Open-Angle Glaucoma
                  Deeplearning Assisted Diagnosis of Primary Open-Angle Glaucoma
                
Accomplishmentsthe solution leverages a finetuned ResNet50 model and Azure Custom Vision Classifier to analyze fundus images for glaucomatous changes
it leverages techniques such as smart tagging for optic disc region identification, suitable for the calculation of cup-to-disc ratio
the datasets leveraged for training and testing include Retina Fundus Images for Glaucoma Analysis (RIGA) and the Dhrishti datasets 

Multivariate Regression and Explainable AI with SHAP
                  Multivariate Regression and Explainable AI with SHAP: exploring factors affecting sleep latency, rem sleep, deep sleep, and number of awakenings.
                
Accomplishmentsdeveloped regression models capable of predicting variables such as awake time, rem sleep time, deep sleep time, sleep latency, and number of awakenings
used SHAP and sensitivity analysis to explain the model's predictions
the models leveraged in this project include Support Vector Machines, XGboost, and TabNet Regressor

Animated Node-link and Adjacency Matrix Transition
                  Animated Node-link and Adjacency Matrix Transition using the Les Miserables dataset
                
AccomplishmentsAn implementation of an animated transition between a force directed graph and an adjacency matrix
Users can hover over a node to enlarge and highlight its direct connections. This will also display the character name and description.
Click and drag nodes to reposition them. Other nodes will repel and move accordingly
After initiating 'Start Transition', interactions are limited to hover details due to overlapping elements that disable other interactions like link highlighting and node dragging
Users can toggle between the node-link and adjacency matrix views
If you re-order nodes in the matrix view, ensure you allow the reordering process to complete before switching back to the node view.
Overlapping names in the matrix view can be resolved by completing the reordering process.

Skills

Frameworks

Keras

TensorFlow

PyTorch

FastAI

PyMC

Django

Flask

Languages and Databases

Python

SQL

R

MATLAB

Shell Scripting

Libraries

NumPy

Pandas

OpenCV

scikit-learn

matplotlib

Cloud and Others

AWS

GCP

Azure

Heroku

Git

Education

University of Waterloo

Ontario, Canada

Degree: Master of Applied Science in Systems Design Engineering

Thesis: Remote Medical Diagnosis in Virtual Reality: A Mixed-methods approach to understanding Patients and Physicians’ Perceptions through Thematic Analysis and Regression Discontinuity Design.

Relevant Courseworks:

Selected Topics in Communication and Information Systems: Advanced Topics in Pattern Recognition (SYDE 770)
- Comparative Analysis: Real-World Weighted Cross-Entropy Loss Functions Across Various Activation Functions
InfoViz for AI Explainability (CS 889)
- Interactive Dashboard for Text Label Exploration
Data Structure in Health Informatics (CS 792)
- Depression Detection System with Decision Trees and LIME
Time Series Analysis (SYDE 631)
- Understanding Impact of Greenhouse Gas Emissions on Global Warming with Structural Time Series

Contact

+15482554426

mustaphaunubi@gmail.com

github.com/MustaphaU

linkedin.com/in/mustaphaunubi