I’m a computer science student focused on deep learning research. I work on language and multimodal models,
with a particular interest in how they learn, adapt, and behave.
Working on Adversarial Neural Cryptography: studying whether neural communication between Alice and Bob holds up against fresh eavesdroppers trained with different architectures.
Comparing five attacker types, including CNNs, Transformers, and a state-space model, while testing how attacker size and training time affect message recovery.
Architected and trained TrooperViz for multimodal safety evaluation across text, image-text, multilingual, and multi-turn inputs, scaling PyTorch training across 8 GPUs and 160,157 examples.
Fine-tuned a vision-language model and merged it with a text-only guardrail using SLERP; achieved 91.83% single-turn and 90.24% multi-turn accuracy on heldout multimodal benchmarks across 21 safety policies, beating tested public guardrails by up to 16.4 percentage points. Works similarly better on Multilingual benchmarks too.
Built data and evaluation pipelines with Pandas and Polars, generating 85,000 synthetic conversations from a 340,473-row data pool and enabling interactive inference through Gradio and vLLM.
Architected an LLM-powered NL-to-SQL pipeline with few-shot prompting, input normalization, 470+ synonym mappings, parameterized SQL sanitization, session memory, and Weights & Biases observability across 5 Azure ML endpoints.
Built an SVD collaborative-filtering recommendation system with scikit-learn and an LLM-based taste profiler, generating updated user preference profiles through automated nightly ETL jobs and achieving 86.7% Precision@5 and 89.8% Recall@10 on a 6-user, 87-recipe benchmark.
Engineered a multi-source recipe ingestion pipeline using Scrapy, social-media APIs, LLM classification, language filtering, and Faster-Whisper transcription to process video and web-based content across 5 source integrations.
Fine-tuned a Qwen 7B model using TensorFlow, Hugging Face, and Lambda, applying LoRA to reduce training cost and memory usage while delivering a domain-specific question-answering model.
Evaluated model configurations across multiple hyperparameters through systematic experimentation, tracking training and evaluation metrics with Weights & Biases to optimize model performance.
Developed an AI outreach agent using LangChain, Exa.ai, and LLM APIs to automate messaging workflows, reducing manual outreach time and improving response rates while documenting processes in Notion.