Resume

Built a real-time rendering engine from scratch in Odin targeting Vulkan 1.3. Designed for low-latency, GPU-bound workloads with full control over the graphics pipeline.

• 10x frame time reduction: Profiled a memory-bound main render pass with RenderDoc and Nsight, restructured GPU-side data layouts to optimize cache coherency and bandwidth utilization, reducing pass time from 5ms to 500µs.
• Bindless resource management: Implemented descriptor indexing for textures and buffers, eliminating per-draw descriptor set updates and reducing CPU-side overhead.
• GPU-driven rendering: Compute shader frustum culling with indirect draw commands, moving scene traversal entirely to the GPU.
• Memory optimization: Fine-tuned VMA allocation strategies, buffer packing, and memory type selection for optimal GPU memory bandwidth on discrete and integrated hardware.
• Mixed-precision pipeline: FP16 render targets and storage buffers where precision allows, KTX2 supercompressed textures with GPU-side format transcoding.
• Profiling infrastructure: Integrated Tracy with GPU timeline zones for frame-level CPU/GPU profiling, enabling systematic identification of pipeline stalls and bandwidth bottlenecks.

Forward+ renderer built on a custom C++ application framework with NVRHI, abstracting over Vulkan, D3D12, and D3D11.

• Compute-driven pipeline: GPU-based frustum culling via compute shaders, with indirect dispatch for light tile assignment and draw call generation.
• Global illumination: Implemented radiance cascade algorithm for real-time indirect lighting, balancing quality and performance through cascade resolution tuning.
• Shader architecture: Written primarily in Slang (NVIDIA's modern GPU shading language) with GLSL, leveraging Slang's module system and generics for maintainable shader code.
• BRDF model: Energy-preserving Oren-Nayar diffuse with robust post-processing chain (bloom, tone mapping).

Framework for latency-critical applications with an embedded Daslang scripting runtime (outperforms LuaJIT, achieves near-native C++ speed). Coupled with NVRHI for a low-level, multi-API rendering abstraction.

Implemented a novel algorithm in Rust projecting inter-frame pixel deltas into a voxel grid, aggregating across multiple calibrated cameras for real-time 3D position reconstruction of moving objects. Compute-intensive pipeline optimized for throughput on CPU with data-oriented design.

• Performance & observability: Built Dynatrace monitoring dashboards for production health metrics, enabling proactive bottleneck identification and reducing system downtime by 40%.
• API modernization: Migrated legacy MuleSoft APIs to Java Spring Boot, improving uptime by 35% and accelerating incident resolution.
• Incident response: Led root cause analysis for production issues, implementing systematic fixes that reduced recurring incidents by 75%.
• Internal tooling: Built admin tools for production data operations, saving ~20 hours/week in support resources.

• ML model training: Trained predictive models for agricultural yield forecasting using Python and scikit-learn, achieving 92% prediction accuracy on crop yield data.
• Performance optimization: Executed a major codebase refactor improving application performance by 75%, reducing technical debt by 40% through modular architecture and optimized queries.
• Scale: Architected CoverCress, a platform serving 5,000+ users, reducing data processing time by 60% through scalable data modeling.
• Data pipelines: Built ETL pipelines in Python and PostgreSQL for multi-source data ingestion, implementing automated validation that cut manual data entry by 80%.

Linear Algebra, Abstract Algebra, Analysis 1–2, Probability Theory, Machine Learning, Data Structures, Principles of Programming Languages, Calculus 1–3