• Built Gemini-prompt-based detection workflows for engagement-bait classification, designing prompts, evaluation criteria, and analysis loops to improve classification precision and recall for policy-aligned content quality decisions and feed-quality ranking applications.

• Developed LLM-based entertaining-content classifiers for under-18 and new-user cohorts, using prompt engineering, rubric design, and offline evaluation to identify high-quality content for recommendation flows.

• Designed and built a recurring new-user holdout experiment in partnership with Data Science to measure the cumulative impact of content quality launches over time, creating a scalable framework for holistic launch evaluation.

• Helped build a centralized content-visibility decision system that unified moderation, quality, and enforcement logic across content surfaces; integrated creator profile signals, reputation scores, moderation labels, and aggregated content quality scores into production decision flows.

• Led and executed the design of Snap's first machine learning model to detect clickbait, achieving a 50% reduction in clickbait views as the technical lead on the project.

• Optimized feed content quality, reducing the hide rate of stories by 10%, and contributing to a 4% increase in new user retention through enhanced filtering mechanisms.

• Contributed to legacy-system deprecation and infrastructure migration efforts that reduced cost and tech debt, including work that delivered approximately $1,000/day in savings.

• Expanded quality and safety enforcement for the first time from video to image-based content to reduce exposure to inappropriate content, and launched opt-in handling for suggestive content in Discover, reducing suggestive prevalence and negative user feedback.

• Built and maintained the first monitoring and alert system for content quality, along with reusable dashboards and runbooks for demotion systems; improving detection of abnormalities, regression response time, and operational visibility; supported oncall investigations for production incidents, ranking anomalies, and content view regressions.

1- Lead researcher in a project with the aim of increasing the resistance of neural network against different adversarial attacks. With the help of auto-encoders, first adversarial samples were detected. Then, some images were reconstructed noiseless. Finally, using the random nullification techniques before training the network, increased the accuracy of classification.

2- Working on the Information Processing Factory (IPF) project, a joint project among the University of California, Irvine, Technische Universität München, and Technische Universität Braunschweig. I am currently working on detecting the emergent behaviors of the system and trying to react safe against these behaviors. By proactively detecting emergent behaviors, the optimal goal is achieving a system with a safe runtime execution.

3- Working on a project with the aim of monitoring the amount of nutrients such as alcohol, glucose, etc. in the mouth through the embedded sensor in the tooth. My recent contribution to this project is that the movement of the sensor can affect these measurements. A pre-trained CNN model that was fine-tuned by the LSTM network is proposed to track the position of the in-mouth sensor by around 1-millimeter error which was published in IEEE TCAD.

Worked on recognizing the most efficient way of AES cryptography by measuring powers in different AES architectures. The leakage for each method was found with the T-test exam. The project was designed and implemented on Sakura-X board with the help of PicoScope and MATLAB, the leakage was found for each of them. The results led to a paper submitted to RTEST 2018.