Featured Projects
AI, ML, and NLP projects — from language models to multimodal retrieval systems.
Uzbek-English Pretrained Language Model (140M Parameters)
Collected a diverse Uzbek-English dataset and trained a BPE tokenizer from scratch (~62,016 tokens). Built a decoder-only modified Transformer with 140.7M parameters and 1024-token context window. Pre-trained on 2x NVIDIA A100 GPUs for 16.5 hours, reaching ~3.5% cross-entropy loss. Planning instruction fine-tuning and downstream alignment.
LLMs from Scratch — GPT-2 Implementation & Fine-Tuning
Full GPT-2 implementation from scratch: text preprocessing, positional encodings, multi-head self-attention, pretraining with causal language modeling. Fine-tuned for spam classification and instruction tuning for prompt-following.
Multimodal Video Search & Interaction with RAG
Retrieval-augmented video Q&A integrating state-of-the-art retrieval techniques for natural language video interaction.
Neural Conversational Chatbot
Seq2seq chatbot with bidirectional GRU encoder and Luong global attention, trained on 220K+ Cornell Movie-Dialogs pairs.