I am a Computer Science student focused on machine learning from first principles, with a specific research interest in reasoning capabilities, mathematical logic, and transformer interpretability.

My approach is dual-track: I build architectures from scratch to understand the math, and I leverage state-of-the-art open weights to push the boundaries of reasoning.

Focus: Reasoning Alignment & Mathematical Logic A specialized exploration into reasoning behaviors, building upon Qwen2.5-Math-1.5B-Instruct.

• Foundation: Leverages the Qwen2.5-Math architecture to study how small-to-mid-sized models handle complex logical chains.
• Methodology: Focuses on optimizing Chain-of-Thought (CoT) performance and step-by-step verification.
• Goal: To understand the limits of reasoning in <2B parameter models and improve logical consistency in mathematical tasks.

Focus: Architecture & Training Dynamics A ground-up implementation of the GPT-2 architecture (124M) to master the engineering behind LLMs.

• Implemented Causal Self-Attention, Multi-Head Attention, and MLP blocks manually.
• Wrote custom training loops and data loading pipelines without high-level trainer abstractions.
• Why: To move beyond "import transformers" and understand the raw tensor operations that drive language generation.

Focus: The Mathematical Foundation A deep dive into the calculus of AI, based on Andrej Karpathy’s curriculum.

• Built Micrograd (backpropagation engine) from scratch.
• Explored loss surfaces, vanishing gradients, and activation function mechanics.
• Bridged the gap between raw Python/Math and PyTorch tensors.

I believe strong ML engineers must also be strong systems engineers. I build low-level systems to ensure I understand the infrastructure that serves my models.

• HTTP Server (Python) — Handled raw sockets and request parsing to understand the transport layer.
• BitTorrent Client — Implemented peer-to-peer protocols to master distributed data transfer.
• Redis-like Server — Built in-memory data structures to understand efficient state management.

• Neuro-AI Alignment: Exploring new paradigms to align model architectures with biological and neurological processes (e.g., biologically plausible learning rules beyond backprop).
• Embodied AI & Robotics: Applying reasoning models to physical control systems to bridge the gap between high-level logic and real-world actuation.
• Reasoning Priors: How to induce stronger logic and Chain-of-Thought in smaller parameter spaces.

I am actively looking to collaborate on reasoning-centric ML or robotics research.

[(https://www.linkedin.com/in/jisha-rajala/)] • [jisharajala@gmail.com]

Thanks for stopping by ✨