Hi I'm Tejus, welcome to my site.
About me
- Completing MS in CS @ UC San Diego, specializing in Artificial Intelligence.
- Completed BS in Applied Physics @ UC Irvine.
- Passions lie in research-driven, open-ended problems in LLM model behavior, applied ML, computational modeling, and forecasting.
- I also enjoy ML Systems design, high-performance computing, and ML acceleration.
Current work:
- Researching representational alignment between human and LLMs and interpretability of black-box models @ UC San Diego. Learn more about my research here.
Notable Experiences and Projects
-
Quantitative Research Intern @ Evovest - Time-series Diffusion models for equities forecasting.
- Worked on noise schedules, sampling distributions, and CFG to enable model to learn signal from very low SNR data.
- Demonstrated significant improvement in the learned correlation structure and uncertainty forecasts.
- Takeaway: Diffusion models learn relationships which transformers and other neural net architectures don't have ability to learn.
-
Worked on an implementation of a quantized transformer on FPGA for low latency, at UCSD.
- Model classified short-lived particle streams generated by the Large Hadron Collider.
- Accuracy-preserving quantization and fast inference were a focus.
-
Researched physics-informed deep learning methods for volumetric medical imaging at UCI.
- Work accepted by the International Journal of Cardiovascular Imaging.