Scalable Machine Learning Models Deployment in CitiBank* Implemented end-to-end ML lifecycle on Azure ML, covering feature engineering, training, validation, and model deployment.* Deployed scalable machine learning models using IaC tools like Terraform and containerized solutions with Docker and Kubernetes.* Designed CI/CD pipelines with Jenkins and GitHub to automate model training, testing, and deployment processes.* Integrated Azure OpenAI for advanced NLP tasks, enhancing model capabilities with generative AI solutions.* Monitored model performance, managed feedback loops, and implemented model drift detection for reliable predictions over time.

Project: Job Connection Model Deployment for Indeed* Developed a Machine Learning Ops platform for job connection forecasting, leveraging Kubeflow, MLflow, and Kubernetes to predict job views, applications, candidate hires, and application volumes.* Integrated Azure cloud services, including Azure Kubernetes Service (AKS), Azure Container Instances (ACI), Databricks, Azure ML, and Databricks MLflow, for scalable ML pipeline deployment and lifecycle management.* Implemented CI/CD pipelines using Azure DevOps, Pipelines, and GitHub to automate the development, deployment, and monitoring of ML models.* Built serverless APIs with Azure Functions and Azure API Management (APIM) for real-time model inference and seamless integration with job posting platforms.* Automated infrastructure management using Terraform and HCL for provisioning and scaling Azure resources, ensuring reproducibility and cost-efficiency.

Project: Job Connection Model Development for Indeed* Delivered the job connection model to predict if a newly published job will obtain applications, messages, interviews, etc. * This model is built by using the Light GBM model and various feature engineering from job attributes and historical data of employers. * The model is serviced as an online estimator of connections of jobs, and used to do job optimizer.Project: Job Attribute Consistency Checker Model Development for Indeed* Developing an LLM-based job attribute extraction, linking, and strength sensing model. Using Open AI GPT-3 and GPT-4 to verify if a skill, license, or education degree is required by the job, and compare it to the selected job attributes to detect the inconsistency between the job description and qualifications.* The candidate is generated by the embedding-based similarity matching, and verified by the GPT model with prompt engineering.

Project: Job Recommendation System for Upwork* Built a job recommendation system to match the job and freelancers using the LLM model. Using the LLM model to "translate" the freelancer's resume to the job description, I fine-tuned the LLM using the hiring records and use it to calculate the matching score between a resume and a job description. The LLM model achieves a F-score of 80%+.Project: Freelancer Skill to Occupation Model for Upwork* Built a skill categorization model to infer the occupation of a skill inserted by the user. It is based on similarity graph of skills. I built the skill similarity graph from the concurrence of the skills in resumes and job descriptions. The label propagation algorithm is used to infer the occupation of a skill from labeled nodes to unlabeled nodes.

Outsourced to Teamflow* Delivered a LLM model to automatically fill a market research form from a Zoom meeting transcript of a sales outreaching meeting. This model is based on GPT-3 and prompt engineering. It reads the entire content of the meeting transcript and generate the pain points, expectations, next steps, etc for the market research purpose.

* Built a digital twin of a city to monitor the social issues and simulate the impact of policies. The digital twin includes social issue maps, time-series forecasting of social issues, and knowledge graphs of social issues. * Delivered the social issue map by integrating the social issue data of substance abuse, divorce, crime, etc.* Delivered AI models to forecast the fertility rate of a city, using time series forecasting model prophet and XGBoost. Features are the historical fertility data of both the city and other coutries/regions. * Delivered AI models to predict the divorces of couples in a city by training a XGBoost model and learning from the couple attributes such as the number of wives, age and marriage history.

Project: Scalable Object Detection Model on AWS SageMaker and EKS* Developed a computer vision project for object detection using AWS services including SageMaker, GroundTruth, and EKS.* Utilized Amazon SageMaker GroundTruth for efficient data labeling to create high-quality training datasets.* Implemented the Single Shot Multibox Detector (SSD) algorithm in SageMaker for robust model training with GPU support.* Deployed the model using SageMaker for hosting with Multi-Model Endpoints, enhancing performance and scalability.* Leveraged Amazon Elastic Kubernetes Service (EKS) for managing scalable deployments with GPUs, ensuring optimal resource utilization.Project: Scalable Face Detection and Cross-Domain Similarity Search System with AWS and PyTorch* Developed a scalable computer vision pipeline for object detection and face similarity search over a 10+ million face database, utilizing Amazon Rekognition for face detection, analysis, and similarity search.* Built and optimized a custom re-ranking model using PyTorch and K-Nearest Neighbors (KNN) on Elasticsearch to match faces across ID photos and natural scenes.* Used AWS GroundTruth for precise labeling of object detection and cross-scene face matching tasks, ensuring high-quality training data.* Deployed the model with SageMaker's Multi-Model Endpoints and GPU support for efficient inference and face similarity search.* Achieved scalable model deployment with Amazon EKS, enabling GPU-backed real-time face detection and matching in natural scenes.

Project: Knowledge Graph* Built a knowledge graph by crawling internet data and a knowledge linking engine. The knowledge graph is used to enrich the internet traffic data. The knowledge graph is built according to an ontology and focuses on organizations and peoples published on internet. The linking is based on the context and entity matching. It uses the NLP model, embedding, and CatBoost model with feature engineering.Project: Realtime Streaming Message Information Extraction* Created a NLP model to read message contents of SMS, WhatsAPP, Telegram, etc to build the message APP user's profiles, extract the events, and build the relation network. It is deployed to analyze the streaming message data and updating the dashboards of profiles, events, and relationship network.

* Developed deep learning model to analyze Arabic text, extract entities, do sentiment analysis, and categorize topics.

* Built machine learning model to predict the interaction between drug and proteins. The model is based on feature engineering with biology knowledge, and a CatBoost model.

* Delivered a ML model to predict the disease from the gene expression data using dimension reduction and classification models.