Local Semantic Search Engine
Brain
Fully local semantic search CLI in Go (brain add, brain search), compiled as a single static binary with no external dependencies.
Combined SQLite + FTS5 full-text search with sqlite-vec vector similarity, using binary quantization to reduce index size. Implemented reciprocal rank fusion to merge keyword and semantic search results. Designed a pluggable Embedder interface.
GoSQLiteFTS5sqlite-vec
Cloud-Based Disposable Workspace Platform
BootLab
SaaS platform to launch temporary, containerized workspaces directly in the browser with no setup or cleanup.
Frontend: React.js + Netlify | Backend: Node.js + AWS EC2 + Docker
React.jsNode.jsAWS EC2Docker
GenAI-Powered Tax Filing Assistant
TaxMan
Agentic RAG application using Gemini, LangChain, and LangGraph for stateful multi-turn conversational workflows.
Function calling, RAG, and agent workflows to parse inputs, apply deductions, and retrieve from official tax law documents. Google GenAI Embeddings + ChromaDB for grounded retrieval.
GeminiLangChainLangGraphChromaDB
Research Study
Convolutional Neural Network
CNN built from scratch in Python (convolutions, pooling, backpropagation), trained on CIFAR-10.
Improved with BatchNorm, Dropout, L2 Regularization, and Optuna. Test accuracy from ~69% → ~82.5%, overfitting gap from ~29% → ~17%. Deployed on Hugging Face Spaces.
PythonCIFAR-10OptunaHugging Face Spaces