About Me
Hey, I'm Pawan. This site's mostly a quiet corner to stash old project docs, ideas I've played with, and the occasional research rabbit hole. Not everything is polished, but it's a good excuse to keep tinkering. Feel free to take a look around :).
Contact
- 📧 pawanpaleja@gmail.com
- đź’Ľ LinkedIn
- đź“Ť Concord, CA, USA
Project Demos
LLVM Compiler Passes: Dataflow Analysis and Program Instrumentation Framework
Implementation of compiler analysis and instrumentation techniques within the LLVM infrastructure. Developed dataflow analysis framework, reaching definitions and liveness analysis, and dynamic profiling tools for instruction counting and branch bias analysis. Follows Sorin Lerner’s CSE 231.
Tag.It
Class project creating a web app forum with efficient data management using Firebase. Led backend development in a project group of 10.
Research Projects
GPU-Accelerated Transformer Framework for Minitorch
A GPU‑optimized transformer system built on top of the Minitorch library. It improves training performance on GPT-2-style decoders using optimizations such as fused softmax and layernorm operators and custom kernels for basic tensor ops. Follows course assignments.
Using Random Restriction to Prove Lower Bounds for Constant-Depth Threshold Circuits
Master’s thesis project applying known mathematical lower bounds for sparse threshold circuits to obtain more efficient learning and compression algorithms.
Learning Implication for Quantum Lower Bounds
Survey paper synthesizing multiple works in recent quantum computing research. Confirms a quantum analogue for converting existing circuit lower bounds into efficient learning algorithms.
Performance guarantees for coloring heuristics on random perfect graphs and generalized split graphs
Donor presentation of new performance guarantees for greedy coloring heuristics on specific graph types.
Cliques to Colors: On Using Clique Statistics to Analyze Heuristic Solvers for the Graph Coloring Problem
An analysis of heuristic solvers’ performance using clique statistics for better predictions and meta-algorithm selection.
Class Projects
Finding Structure in Social Networks
Developed a model for community detection to enhance friendship recommendations in social networks.
Egg-Eater Programming Language
A Rust-based compiler that mimics Python, both basic functions and advanced features such as garbage collection, proper tail calls, and dynamic typing.
Primitive Operations Under Fully Homomorphic Encryption
Survey paper implementing cryptographic primitive operations with the OpenFHE library. Analyzed their performance in comparison to theoretical expectations.