I’m Raphael, an engineering student specializing in data science and machine learning engineering, expected to graduate in 2026.

I enjoy building useful software, designing data pipelines, and integrating modern AI models into real workflows; when a project calls for it, I also like diving into model training and evaluation. I’m especially drawn to interdisciplinary projects and research-oriented problems, where I can study a topic in depth, learn from existing work, and adapt ideas to new constraints.

Here’s my LinkedIn (resume available there):
👉 https://www.linkedin.com/in/raphael-nguyen-162974225/

Below is a selection of my featured projects; feel free to browse the rest of my public repositories.

• Vision API — LLM-Friendly Device Stream & Detection Backend
  https://github.com/Ryustiel/DetectiPublic
  An API designed to be used by LLMs and automation clients: it exposes live visual streams, makes it easy to register or subscribe new devices like phones to live streams, and turns those feeds into structured detection events. Once a stream is available, Grounding DINO can be queried with free-text tags to create detections programmatically.
  Tech: Python, device/stream management, Grounding DINO, event generation.

• Realtime Discord API — Gemini Flash Live-powered Multi-User Conversational Platform
  https://github.com/Ryustiel/MeepiPublic
  Discord-integrated realtime conversational platform built around Gemini Flash Live, so multiple users can participate in the same live conversation. It keeps shared context across participants and channels, supports async MCP-style tool calls, and includes human-in-the-loop (HITL) approval flows for controlled actions.
  Tech: Gemini Flash Live, LangGraph, LangChain, Discord.py, Streamlit, asyncio, Docker Compose, OpenAI, google-generativeai SDK.

• Agents — Stateful Modular Agent Runtime
  https://github.com/Ryustiel/Agenti
  A modular agent runtime for complex, stateful conversations: dynamic and reflective agent runs, long-running tools, callbacks, approvals, context ledgers, prompt-prefix caching, durable memory, structured stubs, external views, live input streams, and OpenAI-compatible access. It separates agent templates from conversation-owned runtime state so each subsystem can evolve independently and be composed into richer agent workflows.
  Tech: Python, FastAPI, PostgreSQL, pgvector, SQLAlchemy, Pydantic, SSE, WebSocket, Docker Compose, OpenAI-compatible API, async workflows, modular LLM orchestration, tool calling, callback-driven tasks, context/prompt caching

• Meep — Multi‑Interface Conversational Platform
  https://github.com/Ryustiel/MeepPublic
  Modular agent built on LangGraph with unified state across multiple chats, async MCP‑style tools, and HITL approvals; designed for scalable, multi‑channel conversations.
  Tech: LangGraph, LangChain, Discord.py, Streamlit, asyncio, Docker Compose, OpenAI, google-generativeai SDK.

• RAGraph — Relational Accessor Graph
  https://github.com/Ryustiel/Ragraph
  Dynamic, multi‑layer knowledge graph for RAG that retrieves context via weighted semantic “accessors” over PostgreSQL + pgvector—great for messy, LLM‑extracted knowledge bases.
  Tech: Python, SQLAlchemy, PostgreSQL/pgvector, Pydantic, OpenAI embeddings.

• Automated Email & Document Processing Pipeline (LLM Storage Service)
  https://github.com/Ryustiel/LLMStorageServicePublic
  Async, containerized pipeline that ingests emails, OCRs attachments, and summarizes content with GPT; pluggable content‑addressable storage (Local, S3, Google Drive).
  Tech: FastAPI, Docker Compose, LangChain, OpenAI, Mistral OCR, S3/GDrive, asyncio.

• Spark + PyTorch: Distributed LSTM for Time‑Series Forecasting
  https://github.com/Ryustiel/SparkTorchPublic
  End‑to‑end workflow using Spark for large‑scale preprocessing and PyTorch DDP via TorchDistributor for distributed LSTM training on time‑series data.
  Tech: Apache Spark, PyTorch (DDP), TorchDistributor, PyArrow, Docker.

• Gemini ImageChat
  https://github.com/Ryustiel/GoogleGenAI
  A Streamlit app to chat with Google’s Gemini image model: attach local images and receive generated images in the conversation; outputs are auto‑saved to project folders.
  Tech: Streamlit, google‑generativeai SDK, PIL, GCP service account.

• Aym — Conversational Graph Builder
  https://github.com/Ryustiel/SI28-Aym
  Turn LLM debates into a living knowledge graph with tool‑proposed node/edge edits and human confirmation, rendered live in an interactive 3D Plotly view; persisted as a single JSON graph.
  Tech: Streamlit, Plotly, LangChain, OpenAI, Pydantic, aiofiles.

If anything catches your eye or you’d like to collaborate, feel free to reach out on LinkedIn.⁣⁠⁣​‌‌‌​​‌​​‌‌​​‌​‌​‌‌‌​​‌‌​‌‌‌​​​​​‌​‌‌‌‌‌​​‌‌​​​​​​‌‌‌​​​​‌‌​​‌​‌​​‌‌​​​‌​​‌‌​‌​​​​‌‌​​‌‌​​‌‌​‌‌‌​‌‌​​‌‌​​​‌‌​​​‌​‌‌​​​‌​​‌‌​​​‌​​‌‌​​​‌​​‌‌​​​​‌​​‌‌​​‌​​‌‌​​​‌​​​‌‌​​​‌​​‌‌​​​​​​‌‌​​​​​​‌‌​‌‌​​​‌‌‌​​​​‌‌​​‌​‌​​‌‌​​​‌​​‌‌​‌‌​​‌‌​​​​‌​​‌‌​‌​​​​‌‌​​​​​‌‌​​​‌​​​‌‌​​​​​​‌‌​‌‌​​​‌‌‌​​​​​‌‌‌​​​​​‌‌​​​‌​​‌‌‌​​‌​‌‌​​‌​‌​​‌‌‌​​‌​‌‌​​‌​​​​‌‌‌​​​​​‌‌‌​​​​‌‌​​​‌‌​‌‌​​‌​‌​‌‌​​‌​‌​​‌‌​‌​‌​‌‌​​​​‌​​‌‌​‌​​​​‌‌‌​​‌​​‌‌​‌​​​​‌‌​​‌‌​​‌‌‌​​​​‌‌​​​‌​​​‌‌​​​‌