CanvasLytics
CanvasLytics [Live | GitHub]
Natural-language data exploration and baseline modelling on an infinite, Excalidraw-style canvas.
Services :
- Agent Service Live
- Pandas EDA Service Live
- Scikit-learn Lab Service Live
- Frontend Client Service Live
Landing Page
Loading
Titanic Dataset
AI Assistant
Titanic Dataset
Titanic Dataset
Titanic Dataset
Architecture
| Layer | Component | Purpose |
|——-|———–|———|
| Client | React / Next.js SPA | Infinite-canvas UI; communicates over WebSockets for live updates. |
| Agent | mcp-agent | Routes user requests to the appropriate MCP tool and manages multi-step workflows. |
| MCP-EDA | pandas-eda | load_csv, describe, missing_values, plot_hist, plot_corr (Pandas + Seaborn). |
| MCP-ML | sklearn-lab | train_model, evaluate_model, feature_importance (scikit-learn). |
| Storage | Postgres + object store | Persists boards, datasets, and generated charts. |
Key Features
| Group | Functionality | |——-|—————| | Data sourcing | One-click load of the public Titanic CSV and automatic schema detection. | | Exploratory data analysis | Descriptive statistics, correlation matrix, missing-value heat-map, and an agent-generated list of suggested follow-up questions. | | Visual analytics | Movable, resizable charts (histogram, box-plot, scatter, stacked bar) rendered with Vega-Lite; optional hand-drawn overlay to mirror Excalidraw’s sketch style. | | Model prototyping | Wizard that trains a baseline scikit-learn pipeline on a user-selected target, then reports accuracy, ROC-AUC, and feature importance; model artefact can be downloaded as a pickle file. | | Narrative & sharing | Auto-generated text summary of EDA findings and model metrics, plus export to PNG, PDF, or read-only share link; per-session version history. | | Performance | Two MCP servers—one for pandas-based EDA tasks, one for scikit-learn modelling—handle requests concurrently; the front-end can stream partial results so long-running jobs do not freeze the UI. |
Excalidraw-Inspired UX
• Single five-colour palette (neutral greys, two blue accents)
• Consistent 24 px grid and generous whitespace
• Shared component library for buttons, menus, and tool-tips
• Toggle to switch charts between clean vector and sketch mode
Quick Start & Local Development
The most reliable way to run the entire application stack locally is with Docker Compose.
# 1. Clone the repository
git clone [https://github.com/SinghArindam/CanvasLytics.git](https://github.com/SinghArindam/CanvasLytics.git)
cd CanvasLytics
# 2. Build and run all services
docker compose up --build
# 3. Open the application in your browser
# Then open http://localhost:3000
Roadmap
- Multi-dataset support (CSV upload)
- Real-time user collaboration
- Additional ML algorithms (e.g., Gradient Boosting, XGBoost)
- Cloud deployment scripts
Contributing
- Fork the repo and create a feature branch.
- Commit using Conventional Commits.
- Open a pull request; CI will run linting and tests.
- Merged PRs are released via semantic-release.
License
Apache 2.0 License - 2025 Arindam Singh (@SinghArindam)
- Arindam Singh