applied cs lab, making a real impact in academia
ABOUT US
Who we are: Our story
Cyrion Labs was founded to bridge the gap between cutting-edge AI research and real-world applications. Our team of researchers, engineers, and innovators work at the intersection of artificial intelligence, machine learning, and computational sciences to drive ethical and impactful technological advancements.
We focus on computational research for social good, partnering with institutions to create AI-driven solutions that address societal challenges, such as safer internet access for students, bias reduction in AI, and accessibility tools for underserved communities. Our research also extends into natural language processing and generative AI, exploring the frontiers of AI-generated content, language models, and human-computer interaction.
Collaboration is at the core of our mission. We work with universities, independent researchers, and industry partners to produce high-impact studies, with plans to publish in leading conferences such as NeurIPS, CVPR, and ICLR. Our work has already contributed to large-scale projects, including a partnership with a public school district of 66,000 students to develop safer internet access solutions.
PARTNERS & SPONSORS
We're extremely grateful for our partners and sponsors who allow cyrion to continue operating, and continue making an impact.
Our Specialities
What we do
CASE STUDIES
CASE STUDIES
Explore our projects.
Explore our projects.
Vega: A Novel Approach to Detecting Modern Web-Proxies in K12 Environments
This paper presents a comprehensive analysis of inherent vulnerabilities in contemporary web filtering systems that rely on manual review and elementary keyword detection, which permit advanced circumvention through domain manipulation and dynamic proxy deployment. We detail how attackers exploit registration of new domains and subdomain configuration to bypass conventional filters, thereby exposing the limitations of existing solutions that primarily utilize client-side scripts for proxy detection. To counter these challenges, we introduce Vega—a novel, multi-tiered detection framework employing a four-level hierarchy. Level 1 implements rapid scanning for distinct JavaScript signatures characteristic of proxy implementations. Level 2 integrates service worker introspection to monitor and detect anomalous fetch-interception behaviors. Level 3 leverages comprehensive HTML content analysis to identify rewriting artifacts indicative of proxy-mediated content transformation. Finally, Level 4 performs rigorous network traffic scrutiny, incorporating protocol-level analysis to detect distinctive markers of backend communications, such as those conforming to TOMPHTTP and Wisp specifications. Vega operates entirely on the client side and incorporates an adaptive caching mechanism to optimize computational load by eliminating redundant scans. Experimental results demonstrate that Vega significantly enhances detection accuracy and response efficiency compared to traditional filtering methods, offering a robust defense against the evolving landscape of web-based circumvention techniques.
Vega: A Novel Approach to Detecting Modern Web-Proxies in K12 Environments
This paper presents a comprehensive analysis of inherent vulnerabilities in contemporary web filtering systems that rely on manual review and elementary keyword detection, which permit advanced circumvention through domain manipulation and dynamic proxy deployment. We detail how attackers exploit registration of new domains and subdomain configuration to bypass conventional filters, thereby exposing the limitations of existing solutions that primarily utilize client-side scripts for proxy detection. To counter these challenges, we introduce Vega—a novel, multi-tiered detection framework employing a four-level hierarchy. Level 1 implements rapid scanning for distinct JavaScript signatures characteristic of proxy implementations. Level 2 integrates service worker introspection to monitor and detect anomalous fetch-interception behaviors. Level 3 leverages comprehensive HTML content analysis to identify rewriting artifacts indicative of proxy-mediated content transformation. Finally, Level 4 performs rigorous network traffic scrutiny, incorporating protocol-level analysis to detect distinctive markers of backend communications, such as those conforming to TOMPHTTP and Wisp specifications. Vega operates entirely on the client side and incorporates an adaptive caching mechanism to optimize computational load by eliminating redundant scans. Experimental results demonstrate that Vega significantly enhances detection accuracy and response efficiency compared to traditional filtering methods, offering a robust defense against the evolving landscape of web-based circumvention techniques.
Vega: A Novel Approach to Detecting Modern Web-Proxies in K12 Environments
This paper presents a comprehensive analysis of inherent vulnerabilities in contemporary web filtering systems that rely on manual review and elementary keyword detection, which permit advanced circumvention through domain manipulation and dynamic proxy deployment. We detail how attackers exploit registration of new domains and subdomain configuration to bypass conventional filters, thereby exposing the limitations of existing solutions that primarily utilize client-side scripts for proxy detection. To counter these challenges, we introduce Vega—a novel, multi-tiered detection framework employing a four-level hierarchy. Level 1 implements rapid scanning for distinct JavaScript signatures characteristic of proxy implementations. Level 2 integrates service worker introspection to monitor and detect anomalous fetch-interception behaviors. Level 3 leverages comprehensive HTML content analysis to identify rewriting artifacts indicative of proxy-mediated content transformation. Finally, Level 4 performs rigorous network traffic scrutiny, incorporating protocol-level analysis to detect distinctive markers of backend communications, such as those conforming to TOMPHTTP and Wisp specifications. Vega operates entirely on the client side and incorporates an adaptive caching mechanism to optimize computational load by eliminating redundant scans. Experimental results demonstrate that Vega significantly enhances detection accuracy and response efficiency compared to traditional filtering methods, offering a robust defense against the evolving landscape of web-based circumvention techniques.
PsyQ: Automated Discovery of Biomarkers for DSM-5 Disorders
This paper presents an in-depth analysis of Tessa, an Embodied Conversational Agent (ECA) designed to provide structured mental health support through natural dialogue. Tessa employs a multimodal sensor suite to capture and interpret user behaviors in real time and uses advanced therapeutic protocols within a structured intervention framework to address cognitive challenges. Its architecture is built around several core components—ranging from data acquisition and adaptive response generation to dynamic intervention calibration and professional escalation—ensuring seamless transitions to human care when necessary. Operating entirely on-device for enhanced privacy and reduced latency, Tessa shows promising improvements in user engagement and adherence, offering an innovative bridge between AI support and traditional therapy.
PsyQ: Automated Discovery of Biomarkers for DSM-5 Disorders
This paper presents an in-depth analysis of Tessa, an Embodied Conversational Agent (ECA) designed to provide structured mental health support through natural dialogue. Tessa employs a multimodal sensor suite to capture and interpret user behaviors in real time and uses advanced therapeutic protocols within a structured intervention framework to address cognitive challenges. Its architecture is built around several core components—ranging from data acquisition and adaptive response generation to dynamic intervention calibration and professional escalation—ensuring seamless transitions to human care when necessary. Operating entirely on-device for enhanced privacy and reduced latency, Tessa shows promising improvements in user engagement and adherence, offering an innovative bridge between AI support and traditional therapy.
PsyQ: Automated Discovery of Biomarkers for DSM-5 Disorders
This paper presents an in-depth analysis of Tessa, an Embodied Conversational Agent (ECA) designed to provide structured mental health support through natural dialogue. Tessa employs a multimodal sensor suite to capture and interpret user behaviors in real time and uses advanced therapeutic protocols within a structured intervention framework to address cognitive challenges. Its architecture is built around several core components—ranging from data acquisition and adaptive response generation to dynamic intervention calibration and professional escalation—ensuring seamless transitions to human care when necessary. Operating entirely on-device for enhanced privacy and reduced latency, Tessa shows promising improvements in user engagement and adherence, offering an innovative bridge between AI support and traditional therapy.
Ellie: A Scalable AI Agent for Mental Health Screening and Intervention
This paper presents a technical overview of a novel multimodal AI framework for diagnosing mental health conditions. The system integrates various data sources—such as text, voice, facial expressions, and posture—using deep learning techniques to analyze psychological symptoms in a way that aligns with clinical standards. Its architecture is built around key components for data acquisition, feature extraction, multimodal integration, and diagnostic decision-making. By blending advanced sensor integration with optimized deep learning methods, the framework offers a scalable, real-time approach to automated mental health assessments, providing a promising bridge between initial AI screening and professional clinical care.
Ellie: A Scalable AI Agent for Mental Health Screening and Intervention
This paper presents a technical overview of a novel multimodal AI framework for diagnosing mental health conditions. The system integrates various data sources—such as text, voice, facial expressions, and posture—using deep learning techniques to analyze psychological symptoms in a way that aligns with clinical standards. Its architecture is built around key components for data acquisition, feature extraction, multimodal integration, and diagnostic decision-making. By blending advanced sensor integration with optimized deep learning methods, the framework offers a scalable, real-time approach to automated mental health assessments, providing a promising bridge between initial AI screening and professional clinical care.
Ellie: A Scalable AI Agent for Mental Health Screening and Intervention
This paper presents a technical overview of a novel multimodal AI framework for diagnosing mental health conditions. The system integrates various data sources—such as text, voice, facial expressions, and posture—using deep learning techniques to analyze psychological symptoms in a way that aligns with clinical standards. Its architecture is built around key components for data acquisition, feature extraction, multimodal integration, and diagnostic decision-making. By blending advanced sensor integration with optimized deep learning methods, the framework offers a scalable, real-time approach to automated mental health assessments, providing a promising bridge between initial AI screening and professional clinical care.
PLVA: A Novel Privacy Layer Model for Visual-Based Web Agents
The increasing reliance on visual data in web applications has amplified concerns regarding the inadvertent exposure of personally identifiable information (PII). In this work, we introduce PLVA—a novel Privacy Layer Model for Visual-Based Web Agents—designed to automatically detect and obscure privacy-threatening elements (PTEs) in web-based imagery. Our framework commences with the development of a Selenium-powered Python application that navigates simulated environments (e.g., Amazon profiles, login screens) to collect screenshots containing PII. These images are annotated using state-of-the-art Visual Language Models (CogVLM, GPT-4o, Qwen VL, Gemini 2.5) to generate bounding boxes around sensitive content, with annotations stored in structured JSON/CSV formats. Building upon this dataset, we explore both fine-tuning of established detection models (YOLOv8, Segment Anything Model, CLIP) and training custom architectures (Faster R-CNN, MobileNet v2/3, YOLO variants) to enhance the precision of PTE identification. Complementing the detection pipeline, we propose a dynamic masking algorithm capable of applying Gaussian blur, pixelation, and inpainting techniques—employing GAN- or diffusion-based approaches—to obfuscate sensitive areas. The masking process is fully parameterizable, allowing adjustments such as blur strength and the hybrid application of inpainting and blurring based on the type of PTE. We validate it by involving a curated test dataset of approximately 100 diverse images, where each image is evaluated across multiple dimensions, including detection accuracy, latency, and resource utilization. Extensive evaluations (1000 iterations in total) demonstrate that the PLVA framework not only reliably identifies and masks PTEs but also integrates seamlessly with existing visual web agents, maintaining overall system performance.
PLVA: A Novel Privacy Layer Model for Visual-Based Web Agents
The increasing reliance on visual data in web applications has amplified concerns regarding the inadvertent exposure of personally identifiable information (PII). In this work, we introduce PLVA—a novel Privacy Layer Model for Visual-Based Web Agents—designed to automatically detect and obscure privacy-threatening elements (PTEs) in web-based imagery. Our framework commences with the development of a Selenium-powered Python application that navigates simulated environments (e.g., Amazon profiles, login screens) to collect screenshots containing PII. These images are annotated using state-of-the-art Visual Language Models (CogVLM, GPT-4o, Qwen VL, Gemini 2.5) to generate bounding boxes around sensitive content, with annotations stored in structured JSON/CSV formats. Building upon this dataset, we explore both fine-tuning of established detection models (YOLOv8, Segment Anything Model, CLIP) and training custom architectures (Faster R-CNN, MobileNet v2/3, YOLO variants) to enhance the precision of PTE identification. Complementing the detection pipeline, we propose a dynamic masking algorithm capable of applying Gaussian blur, pixelation, and inpainting techniques—employing GAN- or diffusion-based approaches—to obfuscate sensitive areas. The masking process is fully parameterizable, allowing adjustments such as blur strength and the hybrid application of inpainting and blurring based on the type of PTE. We validate it by involving a curated test dataset of approximately 100 diverse images, where each image is evaluated across multiple dimensions, including detection accuracy, latency, and resource utilization. Extensive evaluations (1000 iterations in total) demonstrate that the PLVA framework not only reliably identifies and masks PTEs but also integrates seamlessly with existing visual web agents, maintaining overall system performance.
PLVA: A Novel Privacy Layer Model for Visual-Based Web Agents
The increasing reliance on visual data in web applications has amplified concerns regarding the inadvertent exposure of personally identifiable information (PII). In this work, we introduce PLVA—a novel Privacy Layer Model for Visual-Based Web Agents—designed to automatically detect and obscure privacy-threatening elements (PTEs) in web-based imagery. Our framework commences with the development of a Selenium-powered Python application that navigates simulated environments (e.g., Amazon profiles, login screens) to collect screenshots containing PII. These images are annotated using state-of-the-art Visual Language Models (CogVLM, GPT-4o, Qwen VL, Gemini 2.5) to generate bounding boxes around sensitive content, with annotations stored in structured JSON/CSV formats. Building upon this dataset, we explore both fine-tuning of established detection models (YOLOv8, Segment Anything Model, CLIP) and training custom architectures (Faster R-CNN, MobileNet v2/3, YOLO variants) to enhance the precision of PTE identification. Complementing the detection pipeline, we propose a dynamic masking algorithm capable of applying Gaussian blur, pixelation, and inpainting techniques—employing GAN- or diffusion-based approaches—to obfuscate sensitive areas. The masking process is fully parameterizable, allowing adjustments such as blur strength and the hybrid application of inpainting and blurring based on the type of PTE. We validate it by involving a curated test dataset of approximately 100 diverse images, where each image is evaluated across multiple dimensions, including detection accuracy, latency, and resource utilization. Extensive evaluations (1000 iterations in total) demonstrate that the PLVA framework not only reliably identifies and masks PTEs but also integrates seamlessly with existing visual web agents, maintaining overall system performance.
OUR TEAM
Hear from our world-class team of researchers
Hear from our world-class team of researchers
Hear from our world-class team of researchers
-
I've had a passion for making things every since I was young, through Cyrion Labs, I'm able to harness this passion to make a real impact in the world with my skills
Trisanth Srinivasan
-
I’ve always been interested in how AI can help us understand the world. I focus on research that improves how AI learns, reasons, and adapts to solve real-world challenges for our partners.
Rohan Patel
-
Good research does more than explore possibilities, it solves problems. I believe AI should be built with purpose, and at Cyrion Labs, I do just that.
Arjun Deshmukh
-
Technology can change lives, but too often, the people who need it the most are the last to benefit. Cyrion Labs is about closing that gap. Building and deploying AI solutions that actually reach those who need them.
Santosh Patapati
we're always looking for talented researchers, no matter your background, education, or age
we're always looking for talented researchers, no matter your background, education, or age