At Ericsson India Private Limited, I benchmarked WiMamba, a Mamba-based wireless model, against a Transformer. At a 4×4 patch size, it delivered 10.7× lower latency and used 42.6× less GPU memory.
I also evaluated transfer from Florida to San Diego across five tasks, where WiMamba scored 16.6% above baseline and reduced error on channel estimation and interpolation. Fine-tuning with 75% of parameters frozen retained 98.5% of full fine-tuning performance.
In my sequence-modeling project, I implemented S4D, Mamba, Mamba-2, and Mamba-3 from scratch in PyTorch and reproduced Mamba-3’s parity-tracking result. I also built AIRE to evaluate LLM and RAG systems, with replayable runs, deterministic metrics, and regression flags.
For my Autonomous Repository Engineer project, I built a coding agent that verifies task completion through passing tests. My experience also includes writing LLM training data at Alignerr and data annotation, model evaluation, and preprocessing at micro1.
