Projects
Current
- Meta-(out-of-context)-learning in Large Language models: A mechanistic analysis of why we observe the meta OOC learning phenomena in large language models.
- Mechanistic Interpretability Website: A website to contain all things mech interp.
Previous
-
Detecting and Identifying malicious operations by building a behavioral profile for API usage for different customers. This was a really cool project.
-
Helped write the product design document and worked on implementing a ML-based behavior modeling product to provide advanced API protection for securing cloud applications. Prototyped algorithms for detecting OWASP top 10’s most common cyberattacks such as BUA, BOLA etc in real time.
-
Designed and prototyped an ebpf-based kernel level security product for containerized applications running on azure.
-
Went through the MITRE attack framework’s command and control section, focusing on the Application Layer Protocol and simulating a bunch of well known attacks across FTP, SMTP and DNS. Created a C2C server and a barebones testbed to implement prototypes of these attacks. Performed deep packet inspection on a dataset curated from these simulations + real traffic captured at Akamai Technologies and performed unsupervised learning clustering using logs created with Zeek.
-
Replication and comparison of SOTA algorithms for objection detection and tracking from aerial images and videos (implementing and training singe-stage architecture models (all generations of YOLO, SSD models) and two-stage architecture models (R-CNNs)). Generated ensembles to match SOTA performance on Stanford Drone Dataset, UAV-123 dataset and on UAVDT. Also wrote a survey and review of federated learning techniques for object detection and authentication in UAVs.
-
Understanding expression and regulation of lncRNAs in Sorghum Bicolor: Did a Computational analysis of long non-coding RNAs; created models of subcellular localizations, lnc-RNA-mRNA-mi-RNA interactions and lncRNA-protein interactions. Wrote a codebase from scratch to predict cis/trans locations of interest in the sequence and to identify lncRNA sequences from the corpus.
-
Got selected for Google Summer of Code and worked with Red Hen Labs to design and develop a pipeline for NLP techniques to be applied to closed captions data in 15+ languages. Built language specific pre-processing tools, computed frame and sentiment annotation, sentence splitting, tokenization, POS tagging, dependency parsing, lemmatization etc on the text corpus. Deployed the pipeline inside a singularity container on a High Performance Cluster.
-
Autonomous navigation in vehicles: Implemented navigation algorithms for a kobuki Robot. Wrote sensory fusion algorithms to integrate multiple sensory data streams. Trained a deep neural network to allow autonomous navigation of the Kobuki Robot. Re-engineered a car to do autonomous navigation from just a video feed, using a deep neural network trained on video footage of the car driving around the campus.
-
Designed and implemented the front end for Augment Space - a startup in Adelaide, Australia - providing VR tours for real estate on the market. Learned a lot talking to photographers about what it takes to photograph homes.
-
Rank 12 in Regional Mathematics Olympiad (RMO) (Attended training camp for INMO: Indian National Mathematics Olympiad)
-
(3X) Top 5% in Pre-Regional Mathematics Olympiad (pre-RMO)
-
Top 10% nationwide in the National Standard Examination in Junior Science Olympiad (NSEJS)