Building AI systems from first principles agents, memory infrastructure, inference engines, and the tools underneath them.
Currently focused on learning by building: turning ambitious ideas into working, tested software, one system at a time.
- AI Agent Memory Infrastructure
- LLM Inference Optimization
- Retrieval-Augmented Generation (RAG) Systems
- Multi-Agent Systems
- Backend Engineering & System Design
- Python
- C++
- FastAPI
- REST APIs
- Celery / Redis
- PostgreSQL
- SQLite
- PyTorch
- Transformers (Hugging Face)
- LLM Applications
- RAG Systems
- Agentic Workflows
- LangGraph
- Git / GitHub
- Linux
- Docker
A production-grade, open-source Python library for AI agent memory infrastructure. Includes SQLite and in-memory storage backends, a scoring/decay/reinforcement layer, local embedding-based hybrid semantic retrieval, memory consolidation, a FastAPI service layer, and a Typer CLI backed by 80+ tests.
An experimental speculative decoding engine for LLM inference, built to rigorously measure — not assume — whether the technique speeds up generation at small and medium model scales. Includes a from-scratch KV-cached baseline, a correctness-verified speculative decoding engine (with an equivalence test that caught a real bug before it shipped), and controlled benchmarks across GPT-2 and Qwen2.5 model pairs.
A fully offline RAG orchestration system, FastAPI gateway, Celery/Redis workers, and a local LLM served via Ollama for private, local-first document retrieval and generation. Currently being rebuilt into a production-grade architecture.
An interactive Streamlit tool that deconstructs Byte Pair Encoding tokenization from first principles and contrasts additive vs. rotary (RoPE) positional encoding built to make transformer internals visible and explorable rather than abstract.
An adaptive personal operating system designed to understand goals, manage information, and coordinate intelligent agents.
- Build production-grade AI systems
- Contribute to open source
- Secure a strong internship
- Publish impactful projects
- Develop deep expertise in backend and AI engineering
Build. Break. Understand. Repeat.
I prefer learning through projects, implementation, and experimentation rather than passive consumption.