Skip to content
Thanyapisit Buaprakhong
All projects

Agentic data preparation platform

PrepPilot

An open-source coding agent that turns natural-language requests into reproducible data preparation, AutoML, and analysis workflows.

Visit live project View GitHub organization
Project
Undergraduate thesis and open-source project
Role
AI agent architecture, full-stack engineering, and applied research
Focus
LLM agents, Data preparation, AutoML, Open source
Stack
Next.js 16, React 19, TypeScript, FastAPI, Python, Pandas, scikit-learn, Plotly, LangChain, OpenAI / Anthropic, Prisma, MongoDB, Docker
PrepPilot landing page introducing the AI data preparation platform
The public introduction positions PrepPilot around conversational data preparation and AutoML.
A five-minute walkthrough from imperfect source data to automated cleaning, feature engineering, model training, and prediction.
PrepPilot automatic data-cleaning report showing metrics before and after processing
Auto-clean reports make every applied operation and before-and-after quality metric visible.
PrepPilot configurable feature-engineering setup with per-column controls
The agent proposes a feature pipeline while keeping encoding, scaling, selection, and split controls editable.
PrepPilot feature-engineering result with saved datasets and ranked features
Completed pipelines preserve their steps, output datasets, and feature-ranking evidence.
PrepPilot generated quality assessment with distribution analysis charts
Generated analysis combines a written quality assessment with distributions and actionable recommendations.

The project

PrepPilot is an undergraduate thesis and open-source AI data-science workspace for people who need to prepare and model datasets without hand-writing every pipeline. Users upload data, describe the outcome in conversation, review the agent's plan and editable configuration, then receive cleaned datasets, interactive analysis, trained models, and prediction outputs in one workspace.

  • Provides 417 registered handlers across statistics, cleaning, transformation, visualization, feature engineering, NLP, and analysis.
  • Keeps common operations deterministic while preserving a sandboxed code-generation fallback for requests outside the handler catalog.
  • Carries work from raw dataset inspection through cleaning, feature preparation, model training, evaluation, and prediction in one conversational workspace.

Product flow

How it works

  1. 01

    A user uploads a dataset in one of 20+ supported formats and attaches it to a persistent conversation.

  2. 02

    A lightweight router narrows the request to one or more of seven data-science categories before a focused planner selects from 417 registered handlers.

  3. 03

    The executor runs a deterministic handler when possible, or generates Python in a constrained sandbox with an automatic retry path for custom work.

  4. 04

    Guided workflows clean data, engineer features, train and evaluate models, generate reports, and save reusable datasets or prediction artifacts.

System design

How it was built

01

Built the authenticated workspace in Next.js 16 and React 19 with dataset tabs, conversation history, configurable preparation panels, and interactive Plotly artifacts.

02

Designed a FastAPI multi-agent backend spanning routing, planning, execution, interpretation, critique, and replanning, with OpenAI and Anthropic model support.

03

Implemented AutoML across 27 supervised and unsupervised algorithms with five-fold cross-validation, Optuna tuning, model storage, and a Docker-based installation path.

System source

Project repositories