featurecanvas 0.1.1

Python SDK for FeatureCanvas — the no-code feature engineering studio

⌁ FeatureCanvas

No-code feature engineering studio with live impact scoring, leakage guardrails, and AI-powered suggestions.

Live Demo · API Docs · SDK on PyPI

---

What is FeatureCanvas?

FeatureCanvas is a visual, drag-and-drop feature engineering studio — think KNIME or RapidMiner, but with four things none of those tools offer out of the box:

Feature	KNIME / RapidMiner	FeatureCanvas
Drag-and-drop transform canvas	✅	✅
Live predictive impact score per node	❌	✅
Leakage guardrails (rule-based)	❌	✅
AI Copilot with stat-grounded suggestions	❌	✅
No lock-in — plain pandas/sklearn export	❌	✅
Real DAG branching (siblings isolated)	Partial	✅
Path-scoped leakage (per branch)	❌	✅

---

Features

🎯 Live Predictive Impact Score

Every applied transform node shows mutual information and correlation against your target column in real time — directly on the node face, not buried in a separate view.

🛡️ Leakage Guardrails

Rule-based detection fires on the patterns that silently destroy model performance in production:

• Fit-before-split — scaling/encoding fitted on the full dataset before a train/test split
• Groupby self-inclusion — target encoding where each row can see its own label
• Target-touched transforms — derived features that are direct functions of the target
• Target binned as feature — the target column itself discretised into the feature set

All findings are path-scoped — a violation on Branch A never appears in Branch B's leakage panel.

🤖 AI Feature Copilot (Claude-powered)

Suggest transforms button → Claude profiles your dataset's actual column stats and returns ranked, explainable suggestions constrained to real executable transform keys. No hallucinated column names, no free-text the user has to translate.

🔓 No Lock-in Code Export

Every transform has a matching codegen() function. Select any node → click Code → get a standalone Python script with plain pandas/numpy/sklearn that runs anywhere with zero FeatureCanvas dependency.

🌿 Real DAG Branching

Two branches off the same parent are genuinely independent — their dataframes are resolved by walking the real ancestor chain, not by replaying a global ordered list. Sibling nodes never contaminate each other's column dropdowns, impact scores, or leakage findings.

💾 Session Persistence

Sessions survive backend restarts via Upstash Redis (DataFrames stored as Parquet bytes, node graphs as JSON). Canvas layout and sparklines restored from localStorage on page reload.

---

Tech Stack

Layer	Technology
Frontend	React 18 + Vite + react-flow + Tailwind
Backend	FastAPI + Python 3.11
ML Engine	pandas, scikit-learn, scipy
AI	Anthropic Claude API
Session Store	Upstash Redis (Parquet + JSON)
Deploy	Vercel (frontend) + Render (backend)
SDK	Pure Python, pip install featurecanvas

---

Transforms (21)

Numeric

log · sqrt · standard_scale · minmax_scale · robust_scale · power_transform · clip_outliers · abs_value · binning · column_ratio · column_diff

Categorical

onehot · label_encode · frequency_encode · rare_group

Datetime

datetime_decompose

Relational

groupby_agg

Cleaning

fillna · drop_column · rename_column

---

Python SDK

pip install featurecanvas

from featurecanvas import FeatureCanvas

fc = FeatureCanvas("https://featurecanvas.onrender.com")

# Upload a CSV and set the target
session = fc.upload("train.csv")
session.set_target("churned")

# Build a pipeline
log_node = session.apply("log", column="monthly_income")
scaled = log_node.apply("standard_scale", column="monthly_income_log")

# Branch off an earlier node
sqrt_node = session.apply("sqrt", column="age")  # sibling of log_node

# Inspect
print(scaled.columns())         # ['monthly_income_log_scaled', ...]
print(scaled.leakage())         # leakage findings scoped to this branch
print(scaled.scores())          # MI scores vs target

# Check for leakage before shipping
risky = session.apply("groupby_agg",
    group_column="city", agg_column="churned", agg_func="mean")
if risky.has_leakage("high"):
    print("HIGH leakage detected:", risky.leakage())

# Export clean Python — no FeatureCanvas dependency
print(scaled.code())

---

Local Development

Prerequisites

• Python 3.11+
• Node.js 18+
• Upstash Redis account (free tier works)
• Anthropic API key (for Copilot)

Backend

cd backend
python -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # Mac/Linux
pip install -r requirements.txt

Create backend/.env:

ANTHROPIC_API_KEY=sk-ant-...
UPSTASH_REDIS_URL=rediss://default:...@....upstash.io:6379

uvicorn app.main:app --reload --port 8000

Frontend

cd frontend
npm install
npm run dev

Open http://localhost:5173

Tests

# Backend (44 tests)
cd backend && python -m pytest tests/ -v

# Frontend integration (requires backend running)
cd frontend && npx vitest run --testTimeout=20000

---

Deployment

Service	Config file	Notes
Render (backend)	backend/render.yaml	Set ANTHROPIC_API_KEY, UPSTASH_REDIS_URL, FRONTEND_URL
Vercel (frontend)	frontend/vercel.json	Set VITE_API_BASE_URL=https://your-render-url/api

---

Project Structure

featurecanvas/
├── backend/
│   ├── app/
│   │   ├── engine/          # Core ML engine
│   │   │   ├── transforms.py    # 21 transforms with codegen
│   │   │   ├── leakage.py       # Rule-based leakage detection
│   │   │   ├── impact_scoring.py # MI / correlation scoring
│   │   │   ├── graph.py         # DAG resolution engine
│   │   │   ├── session_store.py # Redis-backed persistence
│   │   │   ├── copilot.py       # Claude AI suggestions
│   │   │   ├── codegen.py       # Python script export
│   │   │   └── profiling.py     # Column statistics
│   │   ├── main.py          # FastAPI endpoints
│   │   └── schemas.py       # Pydantic models
│   ├── tests/               # 44 backend tests
│   ├── requirements.txt
│   └── render.yaml
├── frontend/
│   ├── src/
│   │   ├── App.jsx          # Main canvas + state
│   │   ├── nodes/           # SourceNode, TransformNode, TargetNode
│   │   ├── components/      # Sidebar, panels, LeakagePanel, etc.
│   │   └── api/client.js    # Axios API client
│   └── vercel.json
├── sdk/                     # featurecanvas PyPI package
│   ├── featurecanvas/
│   │   ├── __init__.py
│   │   └── client.py
│   └── pyproject.toml
└── featurecanvas_test_data.csv

---

Built by

G. Preetham Saxon — B.Tech CSE @ VIIT Visakhapatnam · IEEE Student Branch Vice Chairperson · AI Product Engineer

---

License

MIT © G. Preetham Saxon