Tabify.ai — Instant In-Context Tabular Predictions with Local TabPFN

Firdaus — Thu, 22 Oct 2026 04:13:40 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

Meet Tabify.ai — an instantaneous, privacy-first web application designed for zero-training machine learning on tabular data (CSV), powered by Prior Labs' TabPFN (Tabular Prior-data Fitted Network) foundation model.

The Problem It Solves & Who Gets to Be Lazier:

Traditional tabular machine learning is tedious: data scientists, analysts, and software engineers spend hours cleaning data, setting up complex scikit-learn/XGBoost pipelines, writing cross-validation splits, and fine-tuning hyperparameters (learning rates, max depths, estimators) just to test a hypothesis. Furthermore, uploading sensitive CSV files (like financial records, healthcare data, or company KPIs) to cloud AI APIs poses serious privacy and compliance risks.

Tabify.ai saves all that effort:

Zero Training Required: No fitting loops, no epochs, no hyperparameter tuning. TabPFN behaves like an LLM for tables, delivering predictions in a single forward pass (~0.1 to 1.5 seconds).
100% Local & Private: All data processing runs strictly on your own hardware (NVIDIA GPU, Apple Silicon MPS, AMD ROCm/DirectML, or multi-core CPU). Zero bytes leave your machine.
One-Command Setup: With Docker Compose (docker compose up --build), both the FastAPI backend and Vue 3 frontend start instantly.

Data scientists, backend engineers, business analysts, and hobbyists can now get instant ML predictions on their tables without writing a single line of training boilerplate.

Demo

Interactive Local Web Dashboard: Runs locally at http://localhost:5173 (or http://localhost).
Interactive Swagger REST API: http://localhost:8000/docs.
Demo Video / Walkthrough: (You can insert a link to your Loom, YouTube, or GIF recording here)
Sample Datasets Included: 1-click demos for Iris Flowers (Classification), House Prices (Regression), Daily Store Sales (Time-Series Forecasting), and Sensor Anomaly (Unsupervised Outlier Detection).

Code

The entire project is 100% open-source under the Apache-2.0 License:

Mocca130604 / tabify.ai

Tabify.ai ⚡

Local Tabular Foundation Model Dashboard powered by TabPFN (Prior Labs)
Built for the Hacktoberfest Open-Source AI Challenge • Category: Best Use of TabPFN

📖 Table of Contents

🌟 Overview

Tabify.ai is a privacy-first web application designed for instantaneous, zero-training machine learning on tabular data (CSV). Powered by Prior Labs' TabPFN (Tabular Prior-data Fitted Network), it delivers transformer-based in-context foundation predictions with 100% local execution directly on your own hardware (CPU, NVIDIA CUDA, AMD ROCm/DirectML, or Apple Silicon MPS).

No external cloud APIs, no subscription fees, no data leaks, and zero tedious hyperparameter tuning required.

🚀 Key Features

100% Local Inference & Total Data Privacy: Zero telemetry or…

View on GitHub

GitHub Repository: https://github.com/Mocca130604/tabify.ai

How I Built It

Tabify.ai is built around Prior Labs' TabPFN v2 open-weight foundation model:

The AI Core (Prior Labs TabPFN):
- Instead of fitting a model specifically to a dataset, TabPFN has been pre-trained on millions of synthetic tabular tasks to perform in-context inference.
- We integrated both TabPFNClassifier and TabPFNRegressor into our backend engine, supporting 4 operational modes:
  - Classification: Discrete category predictions with calibrated class probabilities.
  - Regression: Continuous numerical predictions with automated R², MAE, and RMSE holdout evaluation.
  - Time-Series Forecasting: Automated chronological lag extraction and rolling feature generation.
  - Unsupervised Anomaly Detection: Multivariate density score evaluation to pinpoint tabular outliers.
Hardware-Aware Compute Layer (PyTorch & CUDA/MPS/DirectML):
- Dynamic device catalog that autodetects NVIDIA GPUs (CUDA), AMD GPUs (ROCm/DirectML), Apple Silicon (MPS), and multi-core CPUs, allowing seamless hardware switching through the UI or REST API.
Backend Service (FastAPI & Uvicorn):
- High-throughput asynchronous endpoints for /api/inspect, /api/predict, and /api/health.
- Streaming multipart CSV validation and memory optimization.
Frontend Architecture (Vue 3, Vite & Tailwind CSS):
- Designed with an eye-pleasing Modern Minimalist / Swiss Aesthetic: soft, low-contrast neutral palette, translucent backdrop-blur sticky navigation, active engine health badge, honest multi-phase progress bar, dynamic Chart.js visualizations, and a searchable, paginated table with 1-click CSV export.
Containerization (Docker & Docker Compose):
- Multi-stage Docker build for Vue 3 + Nginx Alpine reverse proxy paired with a Python 3.11 backend container and persistent volume caching for TabPFN weights (tabify_tabpfn_cache).

Why Does Open Innovation Matter?

Open-source and open-weight AI models like TabPFN democratize capabilities that were previously gatekept by multi-million-dollar cloud services.

Data Sovereignty & Privacy: Closed APIs (e.g., sending spreadsheets to proprietary cloud endpoints) are a dealbreaker for regulated industries like healthcare, banking, and confidential business intelligence. Open innovation allows full in-context foundation models to run air-gapped and 100% on-premise.
Deterministic & Inspectable: With open-weight models, there are no surprise API deprecations, rate limits, paywalls, or silent model drift.
Hardware Freedom: Anyone with a modest laptop CPU, an M-series Mac, or a consumer GPU can run state-of-the-art tabular foundation inference without paying per-token or per-row API fees.

My Agent Session

This project was built and accelerated using Google DeepMind's Antigravity AI Agent pair-programming environment. You can explore the prompt iterations, visual refactoring, Docker orchestration, and hardware-device detection logic in the recorded agent workflow session.

(Optional: Add your DevRelay session link here if applicable)

The Lazy Payoff

Time Saved: From 2 Hours to 3 Seconds.

Before Tabify.ai: You obtain a new CSV → install notebook → import pandas/sklearn → handle missing values → encode categorical variables → choose an algorithm → tune hyperparameters → write boilerplate evaluation code → generate plots. Total time: 45–120 minutes.
With Tabify.ai: Drag-and-drop the CSV file → select the target column → click Run → see predictions, metrics, and charts in ~0.2 to 1.5 seconds. You can immediately export the predictions back to CSV and move on with your day.

Prize Categories

Prior Labs Challenge: Best Use of TabPFN (Primary Track)
General Open Source AI Challenge / Best Local AI Application

DEV Community: Firdaus