Agentic AutoML Platform turns datasets and domain documents into production ML models through LLM-orchestrated pipelines. An agentic core powered by LangGraph and MCP tools handles everything from data exploration to deployment, with human-in-the-loop approval gates at every step.
Seven phases mirror the ML lifecycle. Each phase is driven by an LLM agent that proposes actions, generates code, and validates results while the operator approves or edits before execution.
Ingest datasets (CSV, JSON, XLSX) and domain documents (PDF, DOCX, Markdown) into a project workspace. The LLM agent analyzes uploaded data and creates a structured project plan with recommended preprocessing steps, feature engineering strategies, and modeling approaches.
Automated statistical profiling generates distribution charts, correlation matrices, and missing-value analysis on dataset upload. Column-level statistics, data quality scoring, and interactive visualizations surface patterns and issues before any modeling begins.
Query datasets with natural language or raw SQL. A 4-phase pipeline handles intent classification, query generation, execution, and result formatting. Failed queries trigger automatic repair with error context fed back to the LLM.
The LLM agent analyzes raw data and generates Python preprocessing code in notebook cells. Each transformation is proposed, reviewed, executed, and validated through MCP tool calls. Bounded auto-repair retries failed cells with error context rather than silently producing bad output.
Train models through an interactive notebook workspace. The agent generates training code, executes cells in sandboxed Docker containers, and reports metrics. Persistent kernel state maintains variables across cell executions within a session.
Compare trained models on a leaderboard with automatic champion detection and a natural language filter bar. Run Optuna hyperparameter optimization studies with real-time progress streaming. Analyze model errors with decision tree attribution and explore interpretability with SHAP.
Deploy trained models through a dedicated deployment phase with readiness checks, build orchestration, prediction playgrounds, logs, monitoring, and real-time status updates.
LangGraph Orchestration. A state-machine engine coordinates multi-step ML pipelines through MCP tool calls, with phase-aware routing that selects the right tools for each workflow stage.
RAG with Hybrid Search. Ingest domain documents to ground LLM responses in your data. Combines embedding similarity with keyword search for cited, context-aware answers.
Interactive Notebooks. Monaco editor with Jedi-powered Python completions, hover documentation, and syntax highlighting. WebSocket sync with savepoints for checkpoint/restore. Kernel HTML output rendered in isolated Shadow DOM.
Sandboxed Execution. Docker containers with read-only root filesystem, non-root user, and configurable memory/CPU limits. Jupyter Kernel Gateway maintains Python kernel state across cell executions.
| Layer | Technology |
|---|---|
| Frontend | React 19, Vite, TypeScript, Zustand, shadcn/ui, Radix, Tailwind CSS, Monaco Editor |
| Backend | Express 5, TypeScript, LangGraph, MCP SDK, OpenAI SDK, Zod |
| Database | PostgreSQL 16 (metadata, embeddings, notebooks, workflows) |
| Execution | Docker (Python 3.11, scikit-learn, pandas, numpy, Optuna, SHAP) |
| Testing | Vitest (unit), Playwright (E2E), custom eval runner (NL-to-SQL + RAG) |
Prerequisites: Node.js 22 LTS, Docker
npm run install:all # Install backend + frontend + testing + landing + video dependencies
npm run dev # Boot managed Postgres, run migrations, start dev serversThe dev server starts the backend at localhost:4000 and frontend at localhost:5173.
If npm run dev creates or starts the managed automl-postgres-<port> container, it will stop that container on shutdown. Compatible containers that were already running are left running.
backend/ Express 5 + TypeScript API server
src/routes/ Express routers mounted under /api
src/services/ Domain logic (LLM, notebook, websocket)
src/repositories/ File + DB-backed data stores
migrations/ SQL migration files
frontend/ Vite + React 19 SPA
src/components/ UI components (shadcn/ui + custom)
src/stores/ Zustand state management
src/lib/api/ Typed fetch wrappers
landing/ Astro marketing site + public workspace preview
video/ Remotion-based product and branding videos
scripts/dev/ Dev orchestrator (Docker + migrations + servers)
testing/ Playwright E2E benchmarks + eval runner
docs/ Branding assets, API contracts, design system
| Command | Description |
|---|---|
npm run install:all |
Install backend, frontend, testing, landing, and video dependencies |
npm run audit |
Audit root, backend, frontend, and testing dependencies |
npm run dev |
Start development environment (managed Postgres + migrations + servers) |
npm run build |
Build backend (tsc) + frontend (Vite) |
npm run build:landing |
Build the Astro landing site |
npm run test |
Run backend + frontend Vitest suites |
npm run test:landing |
Run the landing Vitest suite |
npm run lint |
Lint backend, frontend, and video |
npm run lint:landing |
Lint the landing workspace |
npm run db:migrate |
Run pending migrations (idempotent) |
npm run benchmark |
Playwright E2E benchmarks (headless) |
npm run eval |
NL-to-SQL + RAG evaluation suite |
npm run benchmark:api |
API load benchmarking (autocannon) |
npm run vercel:landing:pull:preview |
Pull preview settings for the Vercel landing project |
npm run vercel:landing:pull:production |
Pull production settings for the Vercel landing project |
npm run vercel:landing:build |
Build the landing app with Vercel's Build Output API |
npm run vercel:landing:deploy:preview |
Deploy a preview-safe prebuilt landing artifact to Vercel |
npm run vercel:landing:deploy:production |
Promote a prebuilt landing artifact to Vercel production |
docs/api-contracts.md- Request/response contractsdocs/design-system.md- UI guidelines and component patterns