AILAND, https://artificialintelligenceland.com, is a consulting business in AI+NLP space that provides domain experts to clients who are looking for benefiting from recent advances in this area and especially benefit from their textual/unstructured data.

Hiring and building the first AI/ML team in MuleSoft - a Salesforce company. Leadingprojects in different machine learning areas including Search & Discovery, Ranking, Recommendation, Generative AI such as Code Generation LLMs, Conversational AI, and Time-Series Modeling.- Main tools/technologies: Python, SpaCy, Tensorflow, PyTorch, pandas, numpy, Snowflake, AWS Sagemaker, EMR, S3, EC2, Redshift.

- Enhanced NER: Augmented our NER pipeline with patterns to increase precision by 6% and recall by 8%.- Relevant Entity Recognition: Lead the RER model development and deployment based on a BERT-based fine-tuned NER and retrained SpaCy’s NER model (a modified BiLSTM).- News Similarity: an ensemble model that detects whether two articles are similar or not by combining USE, BOW and Random Forest models. This model improved the combined over and under clustering by 25%.- Information Extraction: Lead the development of 6 separate Role/Relation Extraction that finds the action and parties involved in the action in an article, e.g. ”FooCo agreed to acquire BarCo” to (Action: Acquire, Subj: FooCo, Obj: BarCo).- Sentiment: Used RoBERATa within the Transfer-Learning framework to train a model for 3-level sentiment on financial/business related articles.- Multi-Entity Sentiment: trained a BERT model that measures the sentiment for each relevant entity in a given text span. It’s pre-trained on our financial data and then fine-tuned on the multi-entity sentiment downstream task. It achieves average 71% f1-score. Example: ”Morgan Stanley reports that FooCo has filed lawsuit against Barco.” The model detects the sentiment is neutral and negative for FooCo and BarCo respectively.- Key Phrase Extraction: Lead this project to extract key phrases of articles as summary of thematic news. We leveraged a statistical approach together with features such as POS, Name Entities and Dependency Parsing Tree meta-data.- ESG: large scale ensemble multi-label classifiers to detect 41 ESG signals and sub-signals, corporate signals such as M&A, AssetTransactions, etc., and Muni Sectors and Econs, in real time from news articles.- Lead the modeling of a Sentence Boundary Detection model customized to financial news articles.- Entity Linking: a BERT-based EL model whose task is to disambiguate and match the entities found in the article to the entities in our knowledge-base.

- MuniArticles: Trained, tested and deployed to production a model that detects whether an article has important content about municipality bonds. This model saves 7 hours of a human content reviewer per week!- Advertisement Detector: Trained, tested and deployed to production a model that detects whether an article is an advertisement article or not.- Muni Signals: Built and deployed a multi-label text classifier that detects signals regarding municipality bonds in articles’ contents.- Language: Deployed a pre-trained model for language prediction in our production ML micro service.- Signal Predictor: Lead a time-series project for predicting certain signals, such as bankruptcy, happening to a corporation.- Themes: Built and deployed a multi-label text classifier to label articles based on the Themes (e,g. FedsInterestRateCuts, Brexit, TradeWar, et.c) they belong too.- Main tools/technologies: Python(Sklearn, Tensorflow, Keras, Tensorflow-Hub, Pandas, SpaCy, Numpy, etc.), Flask-restful, Docker, Anaconda, MySQL, AWS (S3, EC2, ELB).

- Demographic Models Pipelines: Built multiple pipelines to train/test/evaluate and tune ML models to predict donors’ demographic data such as age and gender and write the predicted data to a Redshift table, scheduled via Apache Airflow.- Donors Analysis: This is a suite of analysis on donors data using EDA techniques in addition to supervised, such as linear regression analysis, and unsupervised, such as k-means clustering, learning which provides insights to donors behavior. This suite included some statistical hypothesis regarding different features of our giving form.- Propensity Score: Implemented, tested, tuned and deployed a set of time-series ML models that predict the propensity to donate for a donor to a specific non-profit organization based on the past behavior of that organization. These models used features from an NLP model that was trained using facebook’s fasttext.- Fraud Detection: Feature engineered and developed an ML model that detects fraudulent credit card transactions in real-time, to protect non-profit organizations from skimmed cards testing attacks. The model was deployed as a REST API on an EC2 instance.- Main tools/technologies: Spark, SparkML, Scala, Java, Airflow, Python(PyTorch, pandas, numpy, flask,..), DeepLearning4Java, PostgreSQL, R, Apache Zeppelin AWS services such as EMR, S3, EC2, Redshift.

- Graph Lookalike: Designed a prototype, trained, tuned, AB-Tested (on 300+ customers), and finally deployed to production, the Graph Based Lookalike product (GB-LAL). It uses features derived from social graph structure together with other features to train a logistic regression-based model that finds users similar to a set of seed users. This has improved the CPI more than 20% for the customers.- LTV Lookalike: Enhanced a Lookalike model based on gradient boosting decision tree with App and Web LTV features; It lifted all the KPIs for the advertisers especially CPI and CPS.- ML Model’s Testing Framework: Designed and implemented an offline testing platform for our LAL machine learning models which facilitates testing different models, such as gradient boosting decision trees(gbdt), random forest or logistic regression, and tuning parameters offline using cross-validation technique.- Main tools/technologies: Spark, SparkML, Scala, Java, Python, Apache Zeppelin, Looker, Google Cloud Services such as Gcs, Dataproc, BigQuery, BigTable, Spanner.

Worked on the models and mechanism to improve the revenue and user/advertiser experience - Ads ranking and pricing of Bing ads

Online Ad-Auctions, Designing Optimal and Sub-optimal algorithms for online e-commerce problems

Network Flow Theory, Deterministic OR Models, Calculus One for Honors Math, Calculus One for Engineering, Linear Programming, Advanced Calculus for ECE students, Discrete Math for ECE students

Combinatorial Algorithms, Graph Algorithms