Resume

Conroe, TX • (936) 828-7896 • raamesh@raameshb.xyz • GitHub

Selected Research Projects

June 2026 – Present

Implemented a configurable Mamba-1 model in JAX and Flax NNX with configurations faithful to both the paper and the official repository; supports exponential Euler and zero-order hold (ZOH) discretization and both broadcasted scalar and learned generalized low-rank $Δ$ selection.
Integrated Mamba’s official CUDA selective scan kernel into JAX via XLA FFI; validated CUDA forward and reverse mode selective scan agreement for exponential Euler and CUDA forward agreement for ZOH, whose reverse mode uses a JAX fallback.
Reproduced Mamba’s induction heads length generalization with 100% accuracy at every tested length through 32,768—the longest evaluated due to compute constraints—after training only at length 256 (128 evaluation samples per length).
On an A100, achieved 1.42× whole-model loss and gradient throughput (D = 64) and up to 1.84× CUDA selective-scan forward throughput at L = 8,192 (B = 8, D = 64, N = 16) versus jax.lax.associative_scan baseline with jax.remat.

2026

Implemented S4 in JAX using bilinear discretization and FFT-based convolution for its parallel form, following the Annotated S4.

github.com/RaameshB/Replicating-S4
Implemented S5 in Equinox using JAX’s parallel associative scan primitive.

June 2026 – July 2026

Built a JAX reinforcement learning pipeline combining behavioral-cloning pretraining with PPO and SVG.
Reimplemented the competition environment as a differentiable JAX simulator to support gradient-based optimization through environment dynamics; trained and submitted an agent.

Oct. 2025 – Dec. 2025

Built an end-to-end PySpark pipeline for distributed ingestion, preprocessing, and topic modeling over multi-terabyte Common Crawl web datasets.

May 2024 – Nov. 2024

Modernized a PyTorch Echo State Network library for Python 3, updating dependencies and CI while repairing training, parameter-registration, and reproducibility issues across 19 commits merged upstream in two pull requests.

Sept. 2024 – May 2025

Designed an asynchronous multi-node hyperparameter-search system in Python: independent CMA-ES workers trained and evaluated Echo State Networks while coordinating trials through a centralized PostgreSQL database; led the lab’s machine learning research and mentored student contributors.

Dec. 2023 – May 2024

Deployed YOLOv8 and RT-DETR through TensorRT on an NVIDIA Jetson Orin Nano and built a real-time pipeline to detect objects and pinpoint their positions in 3D space using Intel RealSense point clouds, RANSAC floor removal, and geometric object-center estimation.

May 2023–Aug. 2023

Prototyped liveness-detection pipelines in Python with OpenCV and MediaPipe for integration into a facial-authentication system.

Aug. 2022 – May 2023

Led control and software development during a team rebuild; prototyped OpenCV vision pipelines in Python, deployed robot code in Java, and created technical training materials for new members.

Languages: Python, Java, SQL, Bash

Machine Learning: JAX ML stack (Flax NNX, Equinox, Optax, Orbax), PyTorch, Hugging Face Transformers, scikit-learn

Models & Methods: Transformers, State Space Models, Mamba, S4, PPO, behavioral cloning, CMA-ES, distributed training, model evaluation

Systems & Data: Linux, Docker, Spark/PySpark, TensorRT, AWS EC2/S3, PostgreSQL, Pandas, Polars

Expected December 2027

B.S. Computational Data Science
GPA: 3.93 • Dean’s List

Relevant Coursework