Jai Agarwal

Key Projects & Responsibilities:

• Led end-to-end development of “Outlier Detection by AI”, an anomaly detection system analyzing customer transactional behavior to flag dynamic anomalies in real-time across both historic and live datasets
• Managed a team of 2 interns, delivering the project within 3 months while mentoring on model architecture, data preprocessing, and evaluation strategies
• Developed tailored Explainable AI modules using SHAP, enabling interpretability for flagged anomalies by identifying hidden feature contributions — improving stakeholder trust and model transparency
• Designed and implemented a GPU-accelerated training and inference pipeline using TensorFlow, Docker, and GCP, reducing model latency by 35%
• Utilized deep learning-based Autoencoders for unsupervised anomaly detection, resulting in a 25% improvement in detection precision compared to rule-based benchmarks

Politically Exposed Persons (PEP) Data Extraction

• Developed a generic data extraction parser to collect Politically Exposed Persons (PEP) profiles from over 500+ national and international official sources, overcoming diverse unstructured web formats and high-security barriers.
• Fine-tuned Microsoft Phi-3 Instruct model to automate the extraction of key information (Name, Designation, DOB, Relatives, etc.). From secured official websites, achieving 97% extraction accuracy.
• Deployed and optimized model inference using vLLMs, reducing response latency by 65% compared to standard inference approaches.
• Built and implemented a deduplication and context-matching algorithm, successfully merging duplicate profiles across multiple sources, and reducing manual analyst effort by approximately 40%.
• Designed and scaled API services to deliver structured PEP data, leveraging MongoDB, Logstash, Elasticsearch, Docker, and Kubernetes, achieving a throughput of 400 requests per second.

Additional Contributions: SEO Strategy & SME Leadership:

• Recognized as a Subject Matter Expert (SME) for bridging technical expertise with business strategy initiatives
• Led SEO strategy and competitive analysis, resulting in a 200% increase in organic leads within 6 months
• Achieved Top 3 Google rankings for 4+ strategic keywords and Top 10 rankings for 10+ keywords

Adverse Media Module

• Developed a sentiment analysis model to classify adverse vs. non-adverse news articles using TF-IDF vectorization combined with Logistic Regression, achieving high classification accuracy for initial article filtering
• Built a custom Named Entity Recognition (NER) model leveraging spaCy pipelines and generalist NER extensions to extract key entities such as Person Names, Organizations, City/State, Amounts, and Article Dates from news content
• Implemented aspect-based sentiment analysis (ABSA) using PyABSA and fine-tuned RoBERTa, BERT, and distilled DeBERTa models to classify sentiment specifically at the entity level
• Developed a relationship extraction model to capture contextual relationships, such as age mentions and associations between entities and geographic locations within news articles using T5 transformer question answering.
• Engineered a generic web parser capable of scraping and processing news text from over 5,000+ diverse sources, automating daily data refreshes within a 24-hour window for real-time updates.
• Stack Used: TF-IDF, Logistic Regression,T5 transformers, spaCy, PyABSA, RoBERTa, BERT, DeBERTa, Python, Web Scraping Libraries (e.g., BeautifulSoup, Scrapy)

Additional Contributions: Event Management Leadership

• Managed and organized the annual organizational event, coordinating logistics, communications, and execution, contributing to the event’s success and enhancing cross-departmental engagement