Skip to content
View YagniPatel's full-sized avatar

Block or report YagniPatel

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
YagniPatel/README.md

Hi, I'm Yagni Patel 👋

AI Engineer & Software Engineer | MSc Computer Science (AI Specialization) @ Western University

I build and break AI systems — from adversarial LLM attacks and calibration research to production-grade cloud infrastructure. I bring 3.5 years of software engineering depth into AI/ML work, which means I care about systems that actually run reliably, not just models that look good on paper.

🔬 What I'm Working On

  • MSc Research at Western University — LLM robustness, calibration, and multimodal learning
  • Exploring the gap between model confidence and actual correctness in small LLMs
  • Building defenses against adversarial jailbreak attacks on open-source models

🛠️ Tech Stack

AI & ML

Python PyTorch TensorFlow HuggingFace Scikit-learn

Cloud & Engineering

AWS Docker GitLab CI/CD Linux

Languages

Python Java C++

🚀 Featured AI Projects

🔐 Multimodal Phishing Webpage Detection

Fine-tuned RoBERTa + ResNet50 late-fusion architecture on 5,000 labeled Phishpedia webpages. Text-only model achieved 97% accuracy (FNR 1.55%). Investigated why multimodal fusion failed to beat the text baseline — identified modality imbalance as the root cause.

🎯 LLM Calibration via Self-Evaluation

Built a self-evaluation framework to extract confidence scores from LLM self-judgment logits on TriviaQA. Applied temperature scaling to reduce ECE from 0.217 → 0.132 (Qwen-2.5-1.5B). Found that stronger discriminative ability (AUROC) doesn't imply better calibration.

🛡️ Breaking and Securing LLMs

Evaluated PAIR, GCG, and Prompt-RS attacks against LLaMA-3.1, LLaMA-4, and Qwen3-32B. Prompt-RS hit 99% ASR on Llama models. Designed a two-stage defense pipeline (prompt sanitization + LlamaGuard) achieving 95% Defense Block Rate across all attack paradigms.

💼 Work Highlights

Software Engineer @ Infor (3.5 years)

  • Led AWS OpenSearch upgrade (v1.2 → v2.17), zero downtime, ~25% IOPS improvement
  • Migrated 500 GB critical data from FSx to S3 with zero customer impact
  • Architected Cloud-to-Cloud producer-consumer system → $400K annual cost savings

📫 Let's Connect

LinkedIn Email

Open to AI Engineer, ML Engineer, and Research Engineer roles.

Pinned Loading

  1. llm-calibration-self-evaluation llm-calibration-self-evaluation Public

    Self-evaluation framework for LLM confidence calibration. Extracts True/False logits on TriviaQA; temperature scaling reduces ECE from 0.217→0.132 (Qwen-2.5-1.5B) without model modification.

    Jupyter Notebook

  2. breaking-securing-llms breaking-securing-llms Public

    Evaluating PAIR, GCG, and Prompt-RS jailbreak attacks against LLaMA-3.1, LLaMA-4, and Qwen3-32B. Two-stage defense pipeline (prompt sanitization + LlamaGuard) achieving 95% Defense Block Rate.

    Jupyter Notebook

  3. job_scheduler_using_reinforcement_learning job_scheduler_using_reinforcement_learning Public

    Job scheduling optimization using reinforcement learning. Agent learns scheduling policies to minimize wait times and maximize resource utilization.

    Python

  4. phishing_webpage_detection_using_multimodal phishing_webpage_detection_using_multimodal Public

    Multimodal phishing detection using fine-tuned RoBERTa (97% accuracy) and ResNet50. Late-fusion architecture with analysis of modality imbalance as the calibration bottleneck.

    Python