Syed Mohsin Ali

I designed and implemented a retrieval-augmented generation (RAG) pipeline for advanced analysis of malicious emails. The system processes a dataset of over 10,000 known malicious emails, which are semantically chunked and stored in a FAISS vector database to enable efficient similarity search. To handle long email content, I applied semantic chunking techniques that preserve contextual relevance during vectorization.

Each email is indexed with its sender address and metadata related to the recipient mailbox, allowing context-rich retrieval. When our custom classifier flags a new malicious email, the system performs a vector-based search to retrieve semantically similar historical samples.

An LLM is then prompted with these retrieved samples to generate a comprehensive analytical summary. The response provides insights such as the recurrence of similar scams, behavioral patterns of the sender (e.g., repeated use of the same email address, tone, or scam structure), and intent detection. This enables the identification of scams that follow a similar template with only minor lexical changes, offering deeper threat intelligence and helping to preempt future attacks.

I designed and developed an advanced credential-stealing classifier to detect sophisticated phishing attempts targeting the theft of login credentials. The solution analyzes email content and attachments—including EML, ZIP, HTML, and PDF formats—extracting over 100 categorical features from email headers, subjects, bodies, and file structures.

To ensure high precision, I retrained a BERT model using fine-tuning techniques such as parameter-efficient fine-tuning (PEFT) and low-rank adaptation (LoRA). These approaches enabled efficient model adaptation on resource-constrained hardware without sacrificing performance.

I also implemented custom data augmentation strategies and conducted iterative A/B testing to maximize detection accuracy. The resulting classifier significantly strengthened the email security pipeline, enhancing defenses against credential theft and phishing-based attacks.

I designed and implemented a comprehensive data normalization pipeline to standardize diverse text data elements, including email addresses, dates, times, URLs, greetings, concluding remarks, alphanumeric text, and digits. This pipeline ensured consistency across various data sources, improving the overall quality and reliability of the data.

To address specific challenges, I developed customized algorithms to identify and normalize complex data patterns accurately. These included handling special cases, such as copyright symbols, which are often difficult to standardize. By incorporating these tailored solutions, I maintained the integrity of the data across different formats and sources.

This initiative significantly optimized data processing workflows, enhancing data quality and more efficient analysis. It streamlined data handling and facilitated faster generation of actionable insights, driving improvements in decision-making processes.