Skip to content
Vasudha PasumarthyLLM Researcher

LLM Researcher · Empirical evaluation & safety

Vasudha Pasumarthy

I run controlled experiments on large language models: how they break under adversarial prompting, what fine-tuning can repair, and how much an evaluation pipeline can be trusted. My work spans adversarial robustness, retrieval-augmented generation and LLM observability.

Start a conversationRead the publications
Figure 1 — ProSense · Mistral-7B

Adversarial reasoning feedback and LoRA fine-tuning, measured before and after.

  • Adversarial failure caseslower is better
    before
    787
    after
    477
  • Successful perturbation detectionshigher is better
    before
    25
    after
    340

Evaluated on AdversarialQA (1,000 examples) and TruthfulQA (817 examples). The numbers are the measured values; each pair of bars is scaled to its own row.

Publications
2 peer-reviewed papers — SEDE 2025 (Springer) and Journal of Information Security.
Education
M.S. Computer Science, AI & Machine Learning — GPA 3.93 / 4.0.
Built
5 applied projects — RAG assistants, LLM observability, adversarial fine-tuning.

01 — Research focus

Questions I keep coming back to

Everything here is empirical: a pipeline, a benchmark and a measurement, rather than an opinion about what a model should do.

  • LLM safety & alignment

    How models behave under adversarial pressure, and whether feedback collected during training actually repairs the behaviour.

  • Model evaluation & experimentation

    Evaluation pipelines and benchmark design, including measurement of hallucinated reasoning and model uncertainty.

  • Adversarial robustness

    Adversarial prompting, perturbation detection and iterative fine-tuning of Mistral-7B (ProSense, SEDE 2025).

  • Retrieval-augmented generation

    Chunking, embeddings and vector search with FAISS and LangChain, answering questions over user-supplied documents.

  • Trustworthy & interpretable systems

    Reliability, uncertainty and interpretability treated as measurable properties of a system rather than intentions.

  • Cognitive-inspired learning

    Parallels between infant cognitive learning and unsupervised representation learning with VAEs and GANs.

02 — Experience

Where the work happened

  1. Research Assistant

    Sep 2025 — May 2026

    University of Southern Mississippi

    • Designed an empirical evaluation pipeline combining LLM-generated sentiment with quantitative financial indicators to investigate stock market forecasting performance.
    • Investigated the effects of hallucinated reasoning and model uncertainty on financial prediction reliability through controlled experiments with large language models.
    • Conducted multimodal machine learning experiments using numerical, textual and visual data to evaluate model performance across heterogeneous data sources.

  2. Graduate Research Assistant

    Sep 2023 — Dec 2023

    University of Southern Mississippi

    • Investigated parallels between infant cognitive learning and modern unsupervised learning methods, exploring representation learning and abstraction in AI systems.
    • Studied Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs), analysing the similarities and differences between computational and human learning.
    • Presented research findings on the opportunities and limitations of unsupervised learning, including adaptability, interpretability and ethical considerations.

  3. Software Developer

    Jan 2023 — Jun 2023

    Lampro Tech LLP, India

    • Developed an OCR-based CAPTCHA recognition pipeline using FastAPI and AWS.
    • Built retrieval-based NLP workflows to improve response quality and information retrieval.
    • Prepared and processed datasets used in machine learning experiments while supporting secure data handling practices.

03 — Publications

Peer-reviewed work

Two papers, both with a measurable result attached: robustness against adversarial prompting, and intrusion detection with ensemble models.

SEDE 2025 · SpringerOct 2025

Defending Text Generating LLMs using Adversarial Feedback

  • Published research on improving the robustness of large language models against adversarial prompting, through adversarial reasoning feedback and fine-tuning techniques.
  • Fine-tuned Mistral-7B using LoRA and Unsloth quantization, incorporating adversarial samples for iterative robustness evaluation.
  • Evaluated on the 1,000-example AdversarialQA and 817-example TruthfulQA benchmarks, reducing adversarial failure cases from 787 to 477 while increasing successful perturbation detections from 25 to 340.
DOI 10.1007/978-3-032-08649-5_20
Journal of Information SecurityMay 2024 — Sep 2024

Optimizing Network Security via Ensemble Learning

  • Published research on network intrusion detection using ensemble machine learning models with Bayesian hyperparameter optimization.
  • Evaluated multiple ensemble learning approaches to improve detection performance while reducing false positives.
DOI 10.4236/jis.2024.154030

04 — Projects

Systems I built to answer a question

  1. ProSense — Adversarial Feedback Defense for LLMs

    Feb — May 2025

    • Proposed the research idea and co-developed the ProSense framework for improving robustness against adversarial prompting.
    • Contributed to the orchestration design, implementation, fine-tuning and evaluation of Mistral-7B using LoRA for adversarial robustness research.
    • Designed evaluation workflows using TruthfulQA to analyse model behaviour and measure improvements in response reliability under adversarial conditions.
    • Mistral-7B
    • LoRA
    • Unsloth
    • TruthfulQA

  2. AgentSleuth — LLM Observability Platform

    Ongoing

    • Developed a Python-based LLM observability pipeline using Qwen2.5-1.5B-Instruct to capture and analyse model inference behaviour.
    • Implemented telemetry collection for inference latency, input/output token counts, request metadata and execution status, with records persisted in PostgreSQL.
    • Designed and executed structured prompt evaluations across five categories, validating telemetry collection and database integration through 30 successful inference requests.
    • Python
    • Qwen2.5-1.5B-Instruct
    • PostgreSQL
    • Telemetry

  3. AI Knowledge Assistant

    Sep 2025

    • Built a retrieval-augmented AI assistant that answers questions over user-provided documents through semantic search.
    • Implemented document chunking, embeddings, FAISS vector search and LangChain orchestration for context-aware response generation.
    • Integrated Gemini 2.5 Flash for response generation and developed a Streamlit interface for document interaction.
    • Gemini 2.5 Flash
    • LangChain
    • FAISS
    • Streamlit

  4. HotTea — Community Messaging Platform

    Sep 2025

    • Co-developed a Firebase-based MVP that enables local businesses to share time-sensitive offers with followers through a single mobile platform.
    • Implemented authentication, user profiles, real-time messaging and Firebase database integration for the application.
    • Collaborated on product planning, technical feasibility and MVP development.
    • Firebase
    • Authentication
    • Real-time messaging

  5. AI Automation Workflows

    Jan 2026

    • Developed automation workflows integrating Gmail, Google Sheets, SerpAPI and LLMs for outreach, structured information extraction and workflow automation.
    • Designed reliable JSON-based processing pipelines with batching and sequential execution strategies to improve robustness and manage API rate limits.
    • n8n
    • APIs
    • JSON pipelines

05 — Toolkit

What I reach for

Research interests

  • Empirical AI Research
  • LLM Safety & Alignment
  • Model Evaluation
  • Adversarial Robustness
  • Trustworthy AI Systems
  • Cognitive-Inspired Learning

And exploring unconventional approaches to understanding and improving AI systems.

AI & machine learning

  • PyTorch
  • TensorFlow / Keras
  • Scikit-learn
  • NLP
  • Transformers
  • Deep Learning
  • Feature Engineering
  • Model Evaluation

LLMs & generative AI

  • Claude
  • Gemini
  • Qwen
  • OpenAI API
  • LangChain
  • RAG
  • Prompt Engineering
  • Hugging Face
  • LoRA
  • LLM Evaluation

Programming & frameworks

  • Python
  • SQL
  • FastAPI
  • Streamlit
  • Git
  • Docker

Data & cloud

  • Pandas
  • NumPy
  • FAISS
  • PostgreSQL
  • AWS (EC2, S3, Batch)
  • Power BI
  • Excel

06 — Education

Study and continued training

  • M.S. in Computer Science — Specialization in AI & Machine Learning

    University of Southern Mississippi

    Aug 2023 — May 2025

    GPA 3.93 / 4.0

  • B.Tech. in Computer Science and Engineering

    MLR Institute of Technology

    May 2019 — Jun 2023

    GPA 3.21 / 4.0

Continued training

  • Loop Engineering with LangChain Deep Agents

    DeepLearning.AI · ongoing

  • Claude 101

    Anthropic

  • AI Fluency: Frameworks & Foundations

    Anthropic

  • AI using Python Nanodegree

    Amazon & Udacity — reinforcement learning through AWS DeepRacer; qualified for the Advanced AI Nanodegree

  • Machine Learning Specialization

    Stanford University & DeepLearning.AI

  • Generative AI Engineering with LLMs Specialization

    DeepLearning.AI

07 — Contact

Let's talk

I'm open to research collaborations and applied LLM research roles — evaluation, robustness, retrieval and observability. Email is the fastest route; the form works just as well.

Direct

Send a message

Research collaboration, a role, or a question about a paper — all welcome.