Currently spending most of the time in Deep learning. Specifically Transformer's models(Encoders, LLM's) training and understanding different architectures.

Global

• CIBMTR - Equity in post-HCT Survival Predictions - 22th / 3,325 (Top 1%) (Silver Medal) (Experiments) (Kaggle) [MAR 2025]
• Stanford RNA 3D Folding - 77th / 1516 (Top 6%) (Bronze Medal) (Kaggle) [SEPT 2025]
• Learning Agency Lab - Automated Essay Scoring 2.0 - 1212th / 2706 (Top 45%) (Experiments) (Kaggle) [APR 2024]
• HMS - Harmful Brain Activity Classification - 1231th / 1206 (Top 45%) (Kaggle) [JAN 2024]
• UBC Ovarian Cancer Subtype Classification and Outlier Detection (UBC-OCEAN) - 1118th / 1326 (Top 85%) (Kaggle) [OCT 2023]
• NeurIPS 2024 - Predict New Medicines with BELKA - 231th / 1336 (Top 12%) (Kaggle) [APR 2024]
• NeurIPS 2023 - Machine Unlearning - 164th / 1336 (Top 14%) (Kaggle) [SEP 2023]

• Quantization Aware Training and Post training quantization using unsloth ang torchAO : (Link)
• Speculative Decoding in vLLM: Complete Guide to Faster LLM Inference : (Link)
  • draft model,n-gram, suffix, mlp speculators, eagle techniques
• The Complete Guide to LLM Quantization with vLLM: Benchmarks & Best Practices : (Link)
  • AWQ, GPTQ, Marlin-AWQ, Marlin-AWQ, GGUF, BnB.
• vLLM Optimization Techniques: 5 Practical Methods to Improve Performance : (Link)
  • Prefix caching, kv-cache quantization, CPU offloading, Disaggregated prefill/decode, zero reload sleep mode.
• Scaling LLM Inference: Data, Pipeline & Tensor Parallelism in vLLM : (Link)
  • Tensor Parallelism, Data Parallelism, Pipeline Parallelism

• Understanding Model Memory Calculations (Link)
• RNA Transformer : Encoder modeling RNA Sequnces to predict 3D Structure (Link)
• Optimizing PyTorch Model Training: Balancing Speed and Memory Efficiency (Link)
• RhoFold+: A Revolutionary Framework for RNA 3D Structure Prediction (Link)

• LLM Reasoning the Emergent Capability (Link).
• GRPO(Group Relative Policy Optimization) (Link).

• AI Agent Frameworks Quick View (Link)
• SHADE Arena : Evaluating Sabotage and Monitoring in AI Agnets. (Link).

• Standing Ovation Award - January 2025 by Wolters Kluwer
• Wolters Kluwer - Code Games 2023 Winner - First runner up

• Experiementation pipeline for image classification. Used to performe 50+ experiments in kaggle competiton named ISIC (Link)
• Experiementation pipeline for text classification. Used to performe 50+ experiments in kaggle competiton named Automated Essay Scoring (Link)
• Developed VLM from scratch in Pytorch (Link)