I work across systems and software engineering — from low-level, performance-critical C++ to applied machine learning and the tooling around it. I'm most interested in problems where correctness and performance both matter: concurrency, memory layout, and software that has to behave predictably under real constraints.
Outside of that, I've been spending more time on embedded systems and robotics fundamentals — RTOS concepts, microcontroller programming, and work that sits closer to hardware.
class AyushmanRaha {
public:
std::string languages[] = { "C++17/20", "Python", "C", "Java" };
std::string focus[] = { "Low-Latency Systems", "Applied Machine Learning",
"Computer Vision", "Embedded Systems (growing focus)" };
std::string currently = "Lock-free C++ · ML pipelines · RTOS & embedded fundamentals";
};|
Lock-free concurrent data structures, cache-conscious memory layouts, and zero-copy I/O — built to avoid the usual overhead of locks and syscalls in throughput-sensitive paths. |
End-to-end ML pipelines — from messy real-world data to trained models to decision-support output, with an emphasis on the engineering around the model, not just the model itself. |
Vision pipelines and local-first desktop applications — model inference, benchmarking, and evaluation tooling that runs entirely on-device. |
Currently building toward: embedded systems (STM32, ESP32), RTOS fundamentals (FreeRTOS), and robotics (ROS 2).
Languages
Systems & Tooling
Machine Learning & Data
App & Backend
Databases
Ultra-low-latency, lock-free messaging engine in C++20
A messaging engine built around a lock-free single-producer/single-consumer ring buffer using C++20 atomics with explicit acquire/release ordering, instead of std::mutex-based queuing. Data structures are cache-line aligned to avoid false sharing, and persistence is handled via a memory-mapped write-ahead log for zero-copy durability. Built with CMake, tested with GoogleTest.
C++20 Lock-Free SPSC mmap
ML pipeline for customer churn prediction
A churn-prediction pipeline with an ETL layer that handles inconsistent real-world CSV schemas, an XGBoost classifier trained with SMOTE to address class imbalance, and a tiered (Low/Medium/High/Critical) risk-scoring output mapped to suggested retention actions. Built with Python and Streamlit.
Python XGBoost Streamlit ETL
Local-first desktop app for monocular depth estimation
A desktop application (Electron + FastAPI + PyTorch/ONNX Runtime) for running and comparing monocular depth estimation models, benchmarking ONNX-accelerated inference, evaluating against ground truth, and exporting 3D point clouds — entirely on-device, with no cloud calls.
Python PyTorch ONNX Runtime Electron
- Lock-free / concurrent data structures in modern C++
- RTOS fundamentals and microcontroller programming (STM32, ESP32)
- ROS 2 basics for robotics and autonomous systems
- Applied ML tooling for real-world, messy data
Open to Software / Embedded Engineering opportunities — let's connect.