I'm a versatile engineer passionate about applying theoretical concepts to real-world challenges, with a diverse background spanning electronics, robotics, and NLP. I thrive on tackling complex problems and rapidly acquiring new knowledge.

• LLM Fine-tuning for Legal Reasoning: Investigating dataset efficiency for fine-tuning smaller language models in legal contexts.
• Agentic Tabular Data Augmentation: Researching the impact of automating tabular data augmentation for deep learning methods.
• [Hobby] Learning Go with NeoVim: Mastering Go fundamentals using NeoVim as an IDE, enhancing my mouse-less workflow.

• AI & Deep Learning: LLM Development (Fine-tuning, Prompt Engineering), Transformer Architectures, NLP, RAG.
• Machine Learning: PyTorch, Hugging Face, Scikit-learn.
• Data & Programming: Python (Expert), Pandas, NumPy, SQL, MATLAB.
• MLOps & Cloud: MLOps (Git, W&B), Docker, AWS (Foundational).
• Tools: Git/GitHub, Linux, Jupyter, VS Code, Streamlit.
• Engineering Specialties: Numerical Optimization, Model Predictive Control, Control System Design.

• Master Thesis: Advanced LLM Fine-Tuning & Pipeline Automation for Legal Reasoning

  • Repository: https://github.com/ReanFernandes/bar-llama
  • Description: Led end-to-end development of LLM solutions, achieving 3-4x performance gains by fine-tuning Llama 2 7B (Q-LoRa) on US Bar Exam questions. Engineered automated data distillation (Llama 3 70B, IRAC method) and evaluation pipelines (Python, PyTorch, Hugging Face Transformers on HPC). Created interactive visualizations (Streamlit, Plotly).
  • Key Technologies: LLM Fine-tuning (Q-LoRa), Data Distillation, Pipeline Automation, Model Evaluation, Prompt Engineering, Python, PyTorch, Hugging Face, HPC, Streamlit, Plotly.

• Auto SLURM Job Handler

  • Repository: https://github.com/ReanFernandes/bar-llama/tree/master/SLURM-related-scripts
  • Description: Developed a Python-based automated job management tool with SQLite for HPC SLURM clusters, demonstrating practical MLOps skills in pipeline automation and resource optimization.
  • Key Technologies: Python, SQLite, SLURM, HPC, Automation, MLOps (Pipeline Automation).

• Non-Linear Model Predictive Control for Crazyflie 2.1

  • Repository: https://github.com/ReanFernandes/flight-control-lab
  • Description: Implemented and comparatively analyzed direct collocation and direct multiple shooting NMPC methods for a nano quadcopter. Involved system modeling, simulation, and performance analysis.
  • Key Technologies: Python, CasADi, Non-Linear Model Predictive Control (NMPC), Control System Design, Numerical Simulation, Data Analysis.

• Optimal Control of a Simulated Furuta Pendulum

  • Repository: https://github.com/ReanFernandes/NOC-project
  • Description: Designed and simulated MPC controllers (Direct Multiple Shooting, Direct Collocation, Real-Time Iteration) using CasADi on MATLAB for swing-up and stabilization of a Furuta pendulum.
  • Key Technologies: MATLAB, CasADi, Model Predictive Control (MPC), Numerical Simulation, Control System Design (ODE/DAE modeling).

• Model Performance Analysis Dashboard

  • Repository: https://github.com/ReanFernandes/model-accuracy-analyser
  • Description: An interactive Streamlit application for visualizing and analyzing model performance results, designed to supplement findings in the paper "A Llama walks into the 'Bar'". It enables analysis of metrics distributions and learning curves from experimental data.
  • Key Technologies: Streamlit, Pandas, Plotly, NumPy, SciPy.
  • Live Demo:

• MCQ Bias Analyzer Dashboard

  • Repository: https://github.com/ReanFernandes/mcq-bias-analyzer-app
  • Description: An interactive Streamlit dashboard for analyzing potential answer selection biases in multiple-choice questions by ML models. It compares predicted vs. ground truth distributions and visualizes label confusion. Used in my paper for the Multi-state Bar Exam to analyze supervised fine-tuning effects.
  • Key Technologies: Streamlit, Python, Pandas, Plotly, NumPy, SciPy.
  • Live Demo: