Projects

All code can be found in this link.

Machine Learning

Diffusion and flow matching | Website | Code

• Implemented sampling loops, classifier-free guidance, SDEdit algorithm, visual anagrams, hybrid images for the Stability AI DeepFloyd IF model
• Trained flow matching model to do class-conditioned MNIST digit generation with UNet architecture, iterative denoising

Chatbot arena | Code

• Using the LMArena dataset, visualized battle distributions with plotly and gradio, explored confounding stylistic variables
• Built model ELO leaderboard like LMArena following Bradley-Terry model

NeRF | Website | Code

• Using 40 images from my camera, trained a NeRF model to render a 3D scene of an object with novel views
• NeRF pipeline from scratch, e.g. volume rendering, sinusoidal positional encoding, ray sampling, PSNR, visualizations with viser

DNABERT | Code

• Trained a classification head on top of pretrained DNABERT-6 representations to perform species identification across human, dog, and chimpanzee genomes
• Implemented preprocessing pipeline using k-mer tokenization (k=6) to convert raw nucleotide sequences into compatible input tokens

CLIP retrieval and captioning | Code

• Performed retrieval and caption generation on Concadia dataset with CLIP embeddings
• Implemented Rational Speech Acts (RSA) e.g. literal listener, pragmatic speaker, pragmatic listener to select the most unambiguous utterance given a set of possible referents

ConvNeXt sound classification | Code

• Used Torchaudio to transform .wav files from UrbanSound8K dataset to 2D spectrograms, then finetuned a ConvNeXt model for sound classification, achieving 90% accuracy
• Compared different 3 training paradigms: no pretraining, pretraining with frozen backbone, and pretraining with full finetuning

Text2SQL finetuning | Code

• Finetuned GPT-2 to generate SQL queries executed against a database and compared performance with few-shot prompting using Hugging Face transformers library

LLM MCQ finetuning | Code

• Finetuned Qwen2.5-0.5B-Instruct using Hugging Face transformers and trl libraries to solve CS189 MCQs with 45% accuracy

LLM tokenization and embeddings | Code

• Implemented BPE tokenization, analyze Zipfian distributions, then trained a Skip-Gram model to compute embeddings using negative sampling
• Used spaCy to learn a linear transformation $W$ that maps two embedding spaces in different languages to perform machine translation, lemmatize and generate dependency trees of text corpus

ResNet-18 | Code

• Full ResNet-18 architecture in PyTorch reaching 70% validation accuracy on a sample of the Imagenet dataset

Autostitching photo mosaics | Website | Code

• Stitching images together to panoramas by estimating homographies, applying projective warping, and blending images with Laplacian pyramid
• Automated mosaic process using Harris corners, ANMS, and RANSAC homography estimation

Language model eval | Code

• Evaluated text/audio classification, machine translation, and open-ended text generation, implemented PRC curves, Levenstein distance, BLEU score, LLM-as-a-judge

Fashion MNIST classifier | Code

• Trained a Fashion MNIST classifier invariant to shifts, rotations, and blurs with scikit-learn

Image filtering | Website | Code

• Core image processing techniques such as convolution, image sharpening, hybrid images, and multi-resolution blending using Gaussian and Laplacian pyramids

Autodifferentiation | Code

• Constructed computation graph of loss functions, implemented downstream gradient propagation using topological sort with Kahn’s algorithm
• Implemented SGD, Momentum, and Adam optimizers

Coloring the Prokudin-Gorskii photo collection | Website | Code

• Aligned three color channels quickly with pyramid search

RISC-V MNIST classifier | Code

• Fully connected neural network with RISC-V Assembly

Software Engineering

Viso | Website | Repo

• Learning platform that transforms math questions into rich animated explanations
• Multi-agent pipeline with smolagents RAG research agent, tool calling, Manim rendering agent, and planning agents
• Won most technical project in AdventureX, Top 20 in amber.ac Hackathon

RISV-C CPU | Code

• 2-stage pipelined CPU in Logisim implementing a subset of the RISC-V ISA

AI Entrepreneurs at Berkeley | Website | Repo

• Founding member of $100M student-run AI incubator, built and maintain the website

Fun

Amateur keyboard player | Recordings

• Demoted from concert grand pianist to digital keyboardist
• Currently sporting a pirated setup with Garritan Yamaha CFX and various transcription tools