projects

  • 2026
    nanoQwen — LLM Implementation from Scratch
    • Implemented a Qwen-style decoder-only Transformer from scratch in PyTorch with RoPE, RMSNorm, Grouped-Query Attention (GQA), causal self-attention, and autoregressive text generation.
    • Built end-to-end pretraining and supervised fine-tuning (SFT) pipelines with custom dataset preprocessing, mixed-precision training, checkpointing, configurable experiment workflows, and tokenizer integration.
    • Developed modular training and inference infrastructure including KV-cache based decoding, HuggingFace parity validation, model evaluation utilities, and configurable scaling for transformer experimentation.
  • 2022
    PrecoderNet — Hybrid Beamforming
    • Implemented deep learning precoder for hybrid beamforming in massive MIMO systems using DDPG, TD3, and SAC reinforcement learning algorithms.
  • 2022
    Model-Free End-to-End Communication Systems
    • Implemented the model-based and model-free auto-encoder based end-to-end communication system for AWGN and Rayleigh Block Fading channels as in: Fayçal Ait Aoudia and Jakob Hoydis, "Model-Free Training of End-to-End Communication Systems."
  • 2022
    Policy Gradient Algorithms for Atari Games
    • Applied A2C, A3C, TRPO, and PPO based on the stablebaselines3 and ray rllib implementations for Pong, Breakout and Space-Invaders atari games.