Akshaya Balaji

I’m a data professional and UC Berkeley alum (M.Eng. EECS, 2022) who spent around three years in the weeds of data analytics—turning messy business requirements into clean, functional data models and SQL solutions.

I love building things, but lately, my curiosity has shifted toward a bigger question: How do we make sure advanced AI systems are actually safe, predictable, and aligned with human intent?

That’s why I’m pivoting my exploration toward AI Safety and Alignment. I’m taking my background in rigorous data validation and systems thinking and applying it to the challenges of interpretability and LLM and agentic AI evaluations. I want to focus on exploring how models can perform reliably for their intended purpose in the real world.

My Tech Stack

• Languages: Python (Pandas, NumPy, Scikit-learn, Matplotlib, Seaborn, PyTorch), SQL
• Machine Learning & Deep Learning: Regression, Classification, Ensemble Methods, Clustering, Model Evaluation, EDA, Neural Networks, CNNs, foundational understanding of Transformers & LLM architectures
• Applied AI: Retrieval Augmented Generation (RAG), Embeddings, Vector Similarity Search, Prompt Engineering
• Data Engineering: Dimensional Modeling, Data Warehousing Concepts
• Tools & Visualization: Docker, Ollama, Git, Jupyter Notebooks

Questions I Plan to Explore

The Evaluation Problem: How can we design hybrid evaluation frameworks for complex LLMs and autonomous agents? Specifically, how do we effectively combine human review with LLM-as-a-judge methods to build scalable, robust benchmarks before these systems interact with the real world?

The Black Box: How can we use mechanistic interpretability to look inside neural architectures, understand how features are represented, and audit a model’s internal reasoning?

Guardrails in Production: What does scalable, real-time monitoring look like for AI systems? How do we build automated telemetry to catch behavioral anomalies, adversarial attacks, and intent drift in the wild?

Safety at the Source: Since data shapes model behavior, how can we leverage rigorous data curation, validation pipelines, and constitutional datasets to bake safety into models from day one?