Amish
Kakka

High-performance, thread-safe in-memory key-value store featuring customizable eviction policies (LRU/LFU), fine-grained synchronization primitives, and low-latency concurrent read/write operations.

Emergency response platform routing community volunteers to victims in real time. Provisioned fault-tolerant AWS backend (EC2 and custom VPC) with Terraform. Achieved sub-second bi-directional GPS tracking using Uber's H3 geospatial indexing and Redis ElastiCache Pub/Sub over WebSockets.

Low-latency C++ library using O(1) rolling-window logic and cache-aware memory layout. Processes 10 years of market data in 0.099 ms — a 25× speedup over NumPy, validated through systematic profiling and benchmarking.

Multi-agent financial research system using Supervisor-Controller architecture in LangGraph to decompose queries across stock screening, sector analysis, and earnings summarization. Live web retrieval grounds outputs and eliminates hallucination. Generates comprehensive reports in under 10 seconds.

Fault-tolerant RAG pipeline on Google Cloud Run. Maintained 1.2s response latency via optimized Pinecone vector indexing and context-window engineering for Gemini 2.5 Flash-Lite.

Fully automated ML lifecycle using a Planner–Validator–Executor agent architecture that converts natural language goals into production-ready Scikit-learn pipelines — handling dataset analysis, feature engineering, model selection, and evaluation end to end.

Fine-tuned Qwen2.5 model on a 15K+ record domain-specific medical dataset using LoRA/PEFT, achieving 90% reduction in trainable parameters. Trained in under 6 hours on Nvidia T4.

Experimental GAN modifications exploring whether replacing raw noise with image embeddings and cross-entropy with content/perceptual loss improves generation quality. Includes a written analysis on Medium documenting findings, failure modes, and open questions.