|
I'm a second-year M.S. student in the Computer Science department at the University of Washington. Currently, my research supervised by Dr. Geetha Thamilarasu focuses on defensive RAG & LLM Security: planning and teaching language models when to ask vs. when to act, with lightweight grounding and prefrence optimization algorithms to choose the best model based on user preference-satisfaction rate.
Additionally, I'm exploring how to overcome memory context-window limitations using short- and long-term retrieval caches, summarization, and Reflexion-style episodic notes that are indexed and reused across tasks.
I'm open to research collaborations. Please email me at [gayat23@uw.edu] if you are interested to explore assistive agents for day-to-day task automation, tool use, and self-evolving large language models.
My Knowledge DB
Projects
LLM Security
A defense framework that guards Retrieval-Augmented Generation pipelines against indirect prompt injection (IPI) attacks. Uses weighted Direct Preference Optimization and constrained RLHF with Lagrangian optimization to train models that distinguish benign retrieved context from adversarially injected payloads, preserving answer quality while rejecting malicious instructions.
A corpus poisoning evaluation framework that stress-tests RAG systems against indirect prompt injection. Generates a poisoned corpus using 13 attack techniques across blackbox, graybox, and whitebox attacker models with semantic alignment via S-PubMedBERT. Provides a standardized benchmark for measuring RAG robustness under realistic adversarial conditions.
High Performance Computing
Optimized a LiDAR-Monocular VO pipeline on KITTI with shared-memory tiling, async transfers, and custom kernels for 2D conv, keyframe selection, and pose estimation improving kernel time by up to 60%.
Computer Vision & Perception
Diffusion-based model predicting 5-second future trajectories in long-tail scenarios using the Waymo Open Dataset; explicitly models multimodal uncertainty.
Panoramic image stitching pipeline that aligns overlapping views and blends them using seam-finding with seamless cloning optimization to minimize visible artifacts at stitch boundaries, producing high-quality composite images.
Natural Language Processing
Text-based classification to detect suicide ideation and attempts using clinical notes and reddit data. After noise removal and abbreviation expansion, compared GloVe, Word2Vec, FastText, and CBOW embeddings. Classifiers with BERT and RoBERTa outperform a Bi-LSTM and SVM baseline, enabling earlier risk identification and ethically informed interventions.
Graph CodeBERT–based detector for zero-day exploits and insecure design flaws (e.g., insufficient compartmentalization and leakage of critical state). It outputs data-flow visualizations and a CLI for triage. Validated on internal repositories mapped to CWE, training combines real code and DCGAN generated synthetic data.
Agents, Planning & Memory Optimization
Building KAG layer memory module to store changing user preferences for long horizon tasks, object affordances, and past interactions, enabling lightweight grounding. Planning is trained with preference optimization algorithms: sequence-level SimPO aligns whole trajectories to user choices, while Step-DPO provides fine-grained <ASK> vs <ACT> feedback.
Machine Learning
Deployed YOLO-based detection to monitor PPE on construction sites in real time (hardhats, vests, goggles). Tracks workers with re-ID, raises on-screen alerts for violations, and logs clips for safety review.
Analyzed pre and post-stroke EEG for hand wave, clench, and face rub motor actions to quantify neural changes and track recovery. Implemented KNN, Decision Tree, SVM, and K-means to classify actions and reveal patterns, informing personalized, data-driven stroke rehabilitation strategies effectively.
Blogs
April Edition — Preference Optimization
Most practitioners pick the highest reward response as chosen and the lowest as rejected. Recent work across seven papers shows this is suboptimal: the absolute worst rejected sample is an outlier that degrades training at scale. I dig into the equations, the evidence, and what it means for building preference datasets.
DPO and SimPO are both offline preference optimization algorithms, but their design choices lead to meaningfully different behavior. SimPO drops the reference model, adds length normalization, and typically scores higher on chat benchmarks. DPO retains a KL constraint that prevents catastrophic drift and better preserves reasoning. I walk through the equations, likelihood displacement, length bias, the math performance tradeoff, and a practical decision framework for choosing between them.
© Gayatri Malladi