Sparse State Space Models for World Modeling
Built a sparse sequence-modeling pipeline using Mamba-2 in JAX with vector quantization for efficient latent representations.
- 4x faster inference through linear-time sequence modeling.
- 60% lower memory usage with discrete latent quantization.
- 25% stronger long-horizon prediction versus recurrent baselines.
World Models
Python
JAX
Mamba-2
Vector Quantization
FakeNewsClassifier Research Package
Designed and trained BERT and RoBERTa classification pipelines with robust benchmarking for noisy and synthetic label settings.
- 95% accuracy across a 50K+ sample training pipeline.
- 12% performance degradation quantified under increasing noise.
- Reproducible robustness benchmarks for comparative model evaluation.
Python
PyTorch
Transformers
BERT
ChatRAG: RAG-WebUI Chatbot
Created a retrieval-augmented assistant with semantic search and vector similarity retrieval over a document collection.
- 40% reduction in hallucinations using retrieval-grounded responses.
- Sub-200ms query latency with optimized embeddings retrieval.
- Semantic search indexing over 1K+ documents.
Python
LangChain
Vector DB
RAG