PyTorch Python CUDA

Built a complete decoder-only transformer language model from first principles. Custom BPE tokenizer, multi-head self-attention, mixed-precision (FP16) training, cosine LR scheduling.

Key Results:

• 38% reduction in training iteration time via FP16
• Trained across 6 dataset configs (1M to 500M tokens)
• Achieved 47.3 validation perplexity on held-out corpus
• Analyzed scaling laws across architecture depth/width

PyTorch State Space Models S6

In-depth study and reproduction of the Mamba architecture, a linear-time sequence model using selective state spaces as an efficient alternative to attention.

Key Results:

• 2.1x throughput improvement over transformer baselines
• 43% reduction in GPU memory at 8K sequence length
• Identified deployment scenarios where SSMs outperform attention (long-context inference, edge devices)
• Maintained research notes repository with experimental results

LangChain Spring Boot RAG Docker

Multi-agent AI productivity platform with RAG pipelines and vector databases for context-aware, personalized assistance.

Key Results:

• 41% latency reduction via async batching & caching
• Sub-180ms response time with quantized 4-bit models
• Deployed on Raspberry Pi 4 (4GB RAM) for edge inference
• End-to-end pipeline: ingestion → embedding → retrieval → inference

LangChain Python FAISS/Chroma Agentic AI

A persistent, full-stack multi-agent system acting as an always-on personal intelligence layer across calendars, tasks, health, finances, and goals.

Key Results:

• Network of specialized sub-agents orchestrated by a master planner LLM
• Long-term memory engine supporting 50,000+ entries
• Sub-50ms retrieval latency
• 65% reduction in manual task intervention via proactive reasoning