Research
Research
Tune in to see more about this project later!!
Built a hybrid vector search system combining high-dimensional image similarity with structured metadata filtering for e-commerce-style retrieval.
Key Techniques include:
Generated 2048-dimensional visual embeddings for 123,000+ product images from the Amazon Berkeley Objects dataset using a pretrained ResNet-50.
Implemented an ACORN-1-style approximate search approach on top of an HNSW index (hnswlib) with metadata-aware query controls, including visit caps and blocked node sets.
Tuned ANN parameters (e.g., M , ef -search, large-k) and evaluated latency–accuracy tradeoffs across different metadata selectivity regimes.
Benchmarked against pre-filter and post-filter baselines using latency (ms) and recall@k.
This project was a group project done by me and the following team members: Ethan Curnow, Miguel Schlicht, Pratyush Sahu.
Designed and implemented a real-time multi-hazard detection system integrating satellite data, environmental sensors, and social media streams for earthquakes, wildfires, and floods.
Key techniques include:
Built an end-to-end data pipeline ingesting NASA FIRMS (VIIRS), USGS earthquake feeds, USGS river gauges, and social media streams (Twitter, Bluesky).
Developed a unified spatiotemporal schema to normalize heterogeneous data sources across location, time, and hazard type.
Engineered multimodal features including satellite-derived indicators, hydrological anomalies, seismic features, linguistic urgency signals, and cross-source correlations.
Implemented a weak supervision framework to generate pseudo-labels using environmental heuristics and multi-source agreement.
Trained hazard-specific Gradient Boosting models and designed a fusion and alerting engine to produce confidence-weighted alerts.
Results: Generated 631 actionable hazard alerts over a 30-day evaluation window
This project was a group project done by me and Ramya Polaki
SnackSense is a real-time snacking detection system using earbuds that sense motion (gyroscope/accelerometer) and sound to passively identify eating behaviors using the built-in motion and acoustic sensors, and a mic. Additionally, a simple ML model was used and trained to detect the type of snack.
The goal of this project is to create a bookstore web application that provides an effortless and seamless browsing experience, allowing users to easily discover and purchase a wide range of books from various genres, authors, and sort by various other options all in one convenient place. With our user-friendly interface, our bookstore web application also provides convenient features such as book reviews and adding a book to a Wishlist, empowering users with valuable insights and information to make informed purchasing decisions, fostering a sense of trust and credibility in our platform. For admin/employees, our application gives a seamless experience for them to add blogs and special deals, add and update books, fulfill orders on time, and a place for employees to see important meetings/events. Lastly, our bookstore web application prioritizes user security and privacy with requiring users to have some form of a complex password and hashing these passwords.
This project has been done in fulfillment of the Capstone Senior Thesis Project as well as the Honors Senior Thesis. This project was a group project done by me and the following team members: William Hobbs, Alfred Lin, Jack Oberman, Sai Oruganti.
The goal of the project was to analyze the data given using k-means clustering to partition the data into an input number of clusters. This project was programmed in C++ using Visual Studio Code IDE.
[Repository]
This project involved the exploration of FPGA-based smart senor designs that incorporate real-time machine-learning for performing online structural state estimation based on vibration signals. I contributed to this project by developing a standalone Foward Pass for LSTM Neural Networks using MATLAB and deploying it onto an FPGA board using SystemVerilog.
This project has been done in fulfillment of USC Honor College Alternative Beyond the Classroom requirement and was advised under Prof. Jason Bakos.
[Repository]
Prototyped and evaluated two course project ideas under Prof. Jason Bakos for a new course called “Internet-of-Things Design” (CSCE 317). The first project idea was to establish communication between an ATmega-based application processor and two axillary processors that control an LCD display and WiFi network interface. The second project idea was a functional weather station that displays sensor data on the attached display and relays the data to network peers.
[Projects created are being used for the USC Course - USC CSCE 317]
Designed, built, and evaluated a system to control temperature inside of a computer enclosure under Prof. Jason Bakos. The system is comprised of an Ardunino Uno microcontroller wired to a temperature sensor, power MOSFET, fan, and 12V power supply. Programmed the microcontroller to read the temperature sensor and control the power MOSFET to govern fan speed using a pulse-width modulated control signal generated by a closed-loop proportional-integral-derivative controller.
[Repository]