Computer Science Student | Systems & ML Infrastructure

gradc is a small, explicit automatic differentiation engine implementing reverse-mode autodiff from scratch. The project focuses on correctness, clarity, and mechanical sympathy rather than convenience-driven abstractions.

The core engine represents computation as an explicit directed acyclic graph, with gradient propagation performed via a topologically ordered backward pass. Memory ownership, graph lifetime, and gradient accumulation are handled deterministically, without hidden allocations or runtime magic.

gradc is intended as:

• a learning vehicle for understanding modern ML systems,
• a foundation for differentiable physics and rendering,
• and a reference implementation that prioritizes transparency over scale.

The design avoids framework-level assumptions and exposes the full computation graph explicitly, making it suitable for experimentation, instrumentation, and extension into domain-specific differentiable systems.

microDR is a small, differentiable rendering engine designed to explore the intersection of graphics, optimization, and learning. It implements a physically motivated forward renderer coupled with a reverse-mode differentiation pipeline that enables gradient-based optimization over scene parameters.

The renderer emphasizes:

• explicit geometry and camera models,
• analytically derived gradients where possible,
• and numerically stable backpropagation through the rendering process.

Rather than targeting photorealism or performance at scale, microDR is built to make the mechanics of differentiable rendering understandable and modifiable. It serves as a testbed for inverse rendering, shape optimization, and vision-from-first-principles experiments.

This project builds directly atop a custom autograd engine, maintaining end-to-end control over graph construction, memory layout, and gradient flow.

knowmore is a knowledge engine designed to support structured, controllable information retrieval across heterogeneous data sources. It combines symbolic indexing techniques with learned representations to balance predictability, interpretability, and recall.

The system treats retrieval as a first-class systems problem:

• explicit data modeling over opaque embeddings,
• deterministic query paths alongside semantic search,
• and stable performance under incremental updates.

knowmore is intended for domains where correctness, traceability, and system behavior matter more than black-box relevance scores. Its architecture supports extension into domain-specific knowledge bases, curriculum-scale content indexing, and long-lived personal or institutional memory systems.

The project emphasizes clean separation between ingestion, indexing, and query execution, enabling experimentation without entangling core system logic.