Projects

Abstract: The rapid expansion of the speech-to-text (STT) tool market has reduced the reliance on proprietary systems. Due to the constant evolution of technology, there exists a continuous need to compare and assess existing market solutions to determine the optimal choice for developers. Our study evaluates self-hosted models and APIs for metrics that are applicable when designing and developing an application that utilizes automatic speech recognition systems. Our data, collected as of mid-March 2025, is up-to-date and highly relevant for developers aiming to build applications leveraging these systems. We chose to evaluate OpenAI’s Whisper, CMUSphinx, Google Cloud STT, Microsoft Azure AI Speech, and Deepgram based on their latency, accuracy, ease of use, and cost-efficiency. Each system transcribes 60 randomly selected hours of audio files from the LibriSpeech dataset. Previous studies only assessed accuracy, which is insufficient when developing a large-scale application that requires the consideration of many other factors. We calculate the latency and average word error rate as well as confirm the advertised cost. Our findings present the best service in each category as well as the best overall.

Timeline: Spring 2025

Keywords: text to speech, speech to text, benchmarking, APIs

Abstract: Code reviewing is the industry standard for how computer scientists and software engineers inspect, critique, and communicate ideas regarding their peers' code. It is an iterative process that is simple to learn. In the context of education, code reviews also offer personalized feedback when teaching students who are new to programming what they should be improving in the code that they write. While the code review process is generally regarded positively, there are some downsides to the system. The first is that leaving effective code reviews requires that the reviewer understands the context of the code that they are reviewing. It is much more difficult for an individual with less programming experience to review an individual with substantial programming experience. The second downside of code reviewing is that it can take a long time to review pull requests and take significant manual effort for comprehensive code reviews. The premise of this project is to determine a way to effectively large language models in saving computer science educators time and effort when reviewing new students' code through code reviews. A Github workflow will be used to output comments provided by the automated reviewer directly to the pull request.

Timeline: Spring 2025

Keywords: code review, automation, generative AI, prompt engineering

Abstract: Searching for specific images within an academic document can be time-consuming and labor-intensive. While manually scrolling through the document to locate an image is the easiest option, it is far from the most efficient solution. However, image processing has several limitations, such as high costs, slow processing speeds, and potential accuracy issues. As a result, there is a need for a more robust PDF image parser to assist researchers. We aim to develop an open-source web application that enables users to quickly identify and locate all images within a PDF document based on specific queries. 

Timeline: Spring 2025

Keywords: image analysis, tagging, PDF

Abstract: This is a team of high performers who will be building out the official mobile ASDRP app for iOS and Android. The app will handle a variety of tasks including form submission (lab safety training, IP training, etc), advisor application/transfer requests, registering for and checking into ASDRP events, and automating the daily sign in process through location tracking. It will deliver real-time program information, including interactive views for application processes, term deadlines, department overviews, NewsBytes updates, and upcoming symposia schedules. Authentication will be done with Google-based sign-in to create personalized profiles for students, mentors, and administrators. The mobile application will enable timely notifications by pushing reminders for key milestones (application deadlines, colloquia, expos). It will track user engagement metrics to guide continuous feature improvements. We are aiming to publish the application in the iOS App Store by around September of 2025. 

Timeline: Summer 2025

Keywords: mobile application development, iOS, Swift, Android, Kotlin, UI/UX design

Abstract: The ASDRP Support Chatbot project delivers a secure, cost-effective, and semantically powerful question-answering system embedded directly into the ASDRP main website (asdrp.org). It leverages a Retrieval-Augmented Generation (RAG) architecture: site content is crawled, chunked, and converted into vector embeddings (via OpenAI’s text-embedding-ada-002), which are stored in a Pinecone vector database. At query time, user questions are authenticated via Google OAuth (restricted to ASDRP domain email accounts), embedded, and used to retrieve the most relevant document passages. Those passages, combined with a concise system prompt, feed into OpenAI’s Chat Completions API (gpt-4-turbo) to generate accurate, context-grounded answers. All interactions are logged in a Wix Data Collection (ChatLogs) for analytics and abuse investigation. User queries will never pollute the core knowledge index. The front end is built with Wix Velo, featuring a dynamic chat widget that appears only after domain-verified login, ensuring both usability and access control. This end-to-end solution balances high-quality semantic retrieval, real-time LLM inference, and stringent security, all within managed or free-tier services to minimize operational costs.

Timeline: Summer 2025

Keywords: Retrieval-Augmented Generation (RAG), customer support, vector embedding, Pinecone, OpenAI, ChatGPT, Wix Velo, Google Cloud, Semantic Search, Content Ingestion & Chunking, Prompt Construction, Prompt Caching, Cost Optimization

**collaboration with Robert Downing Research Group**

Abstract: This project will leverage ASDRP’s own SSH authentication logs to characterize and quantify unauthorized access attempts, identify attacker origins, and inform risk mitigation strategies. By systematically parsing server-side records (/var/log/auth.log), we will extract each login attempt’s timestamp, source IP address, geolocation, and attempted credentials (usernames and passwords). Aggregating these data will reveal attacker distribution by geography, common username/password patterns, and scanning behavior—distinguishing between broad, automated “credential stuffing” campaigns and more targeted probes.  Using statistical analysis and visualization tools, we will measure attack frequency, temporal trends (e.g., peak probing hours), and diversity of source addresses. This evidence-based profiling enables a quantitative risk assessment: estimating the probability of a successful breach under current configurations and projecting potential impact. Building on these insights, we will evaluate and pilot countermeasures—such as adaptive firewall rules (geo-blocking or automated IP blacklisting), enforcement of strong-password policies, rate-limiting, and multi-factor authentication—to lower exposure.

Timeline: Spring 2025

Keywords: Cybersecurity, SSH Authentication, Brute-Force Attacks, Server Log Analysis, Intrusion Detection, Geolocation Tracking, Credential Stuffing, Risk Assessment, Threat Mitigation, Firewall Configuration, Multi-Factor Authentication, Data Analysis 

Abstract: Every semester, our lab faced the recurring challenge of holding students accountable for completing bi-weekly updates in a shared Google Sheet. While the process appeared straightforward, it often resulted in late-night manual reminders, repeated checks of the sheet, and unnecessary back-and-forth messages in Google Space. This manual oversight was inefficient. To address this, we developed an accountability chatbot integrated directly into Google Space, where students were already required to be active. The bot automates reminders, tracks completion, and enforces accountability through a transparent strike system. By embedding the chatbot into the existing communication platform, the system introduces no additional learning curve while ensuring consistent, automated accountability. 

Timeline: Summer 2025

Keywords: Apps Script, Automation, Chatbot, Google Cloud, Python

Abstract: This research project involves training 3 major generative AI models: Midjourney, Dall-E and Stable Diffusion to create a comic, as well as compare their outputs. First, it focuses on using prompt engineering for each AI model. A specific prompt layout has to next be designed by experimenting with ways of phrasing prompts for optimal results. For a comic, it is crucial for the art style and character to be consistent. The individual models using Dall-E, NightCafe's Stable Diffusion, and Midjourney on the art style and characters from different datasets have to be trained. At the end of the fine tuning, each model should be able to adapt to the characters and art style of the sample dataset fed to it. While comparing the outputs, a key finding will be what the best size for such a dataset is. Methods to quantify the outputs for comparison will also be investigated. Lastly, an application that can take a script and dataset of an art style and characters to create a comic will be created. 

Timeline: Summer 2024 - Fall 2024

Keywords: art, comics, generative AI, machine learning, prompt engineering

Contributors: Raina Panda

Abstract: One of the most frequent operations that the ASDRP organization requires is to be able to record attendance of students, parents, faculty, and visitors at events (new student orientations, ASDRP Research Expo) and daily check in at the ASDRP campus. We are developing a proprietary QR Code based web application with Flask, Python, HTML/CSS, Javascript, and deploying it through Render. The application reads from and writes to designated Google Sheets. We designed the system to be able to handle up to thousands of visitors checking in during a short period of time. Our ultimate goal is to fully replace an existing Sign In App subscription for the organization, which costs $500 a year to purchase.

Timeline: Spring 2025 - Spring 2025

Keywords: QR codes, full stack web application development, Python, Flask, HTML/CSS, Javascript, Google Cloud, Docker, Render

Contributors: Tithi Raval

Abstract: At the ASDRP campus, there are multiple deep freezers in the biology lab that need to be monitored at all times. The refrigerators house roughly $1 million worth of cell culture experiments for the Biology and Chemistry departments. If the freezer goes above a certain temperature, this is disastrous for the cells. A previous incident had occurred where the freezer crashed and its alarm went off, but this was at 3AM in the morning when no one was around to hear it and take action. The project proposal is to set up a system that can actively monitor the temperature on the fridge via webcam. An application will be created to take snapshots every 5 minutes and perform image processing using OpenCV and optical character recognition libraries such as EasyOCR. If the temperature is ever above -70 degrees Celsius or absent, the application will call administrator phone numbers until someone picks up to ensure that the notification has not been missed. Such incident monitoring systems already exist for deep freezers, but they are not affordable (in the range of tens of thousands of dollars a year) so there is a need to develop a proprietary solution. This project is high impact because it eliminates manual checking and provides peace of mind in verifying the deep freezers are up and running at any time.  

Timeline: Fall 2024 - Spring 2025

Keywords: incident monitoring, image analysis, automation, optical character recognition

Abstract: Two-Factor Authentication (2FA) significantly strengthens system security by requiring users to provide two independent credentials—something they know (e.g., a password) and something they have (e.g., a one-time code). In this implementation, a one-time password (OTP) is generated and remains valid for 5 minutes by an automated script whenever anyone successfully logs on to the server. The OTP is delivered to the requesting user’s ASDRP email account—applicable to both students and faculty—to ensure that only active organization members can access the server. This project aims to design and deploy a scalable 2FA solution across a Linux server cluster using Pluggable Authentication Modules (PAM) and Python. Leveraging a custom PAM module written in Python, the system will issue and verify time-based OTPs generated and emailed automatically. It also investigates the challenges of constructing a solution that does not rely on PAM.

Timeline: Spring 2025 - Spring 2025

Keywords: Linux, Python, Google Cloud, SMTP, Two Factor Authentication, Multi Factor Authentication