AutoForecast Platform

A self-service AI-powered time series forecasting platform that automates the entire forecasting pipeline — from data profiling to model selection, training, and evaluation.

Architecture

┌─────────────┐     ┌──────────────┐     ┌─────────────────┐
│  React 19   │────▶│  Hono API    │────▶│  SageMaker      │
│  Cloudscape │     │  (Lambda)    │     │  Processing     │
└─────────────┘     └──────┬───────┘     └─────────────────┘
                           │                      │
                    ┌──────▼───────┐       ┌──────▼──────┐
                    │  Bedrock     │       │  AutoGluon  │
                    │  Claude Opus │       │  + Chronos  │
                    └──────────────┘       └─────────────┘

Features

• AI-Driven Analysis — Claude Opus 4.6 analyzes your data and formulates a forecasting hypothesis
• Automated Code Generation — LLM generates custom AutoGluon training code based on data characteristics
• Iterative Validation — Automatic validation loop with up to 3 fix attempts before full training
• Plan-Driven Model Selection — Models chosen by AI analysis, not hardcoded lists
• Real-Time Progress Tracking — 4-step pipeline with elapsed time and status updates
• Interactive Results — Backtest metrics (WAPE), HTML plots, downloadable forecasts

Tech Stack

Layer	Technology
Frontend	React 19, Cloudscape Design System, Vite
API	Hono (TypeScript), esbuild (36KB bundle)
AI	Amazon Bedrock (Claude Opus 4.6)
ML	AutoGluon 1.4 (CPU), Chronos zero-shot
Compute	AWS Lambda (async self-invoke, 15min), SageMaker Processing
Storage	DynamoDB, S3
Infra	CDK, CloudFront, WAF (corp IP restriction)

Pipeline Steps

1. Upload — CSV time series data (up to 1GB)
2. Analysis — AI profiles data, identifies patterns, formulates hypothesis
3. Code Generation — Custom AutoGluon script generated based on analysis
4. Validation — Quick test job (ml.t3.medium) to verify code correctness
5. Full Training — Complete model training with all data
6. Results — WAPE metrics, forecast plots, downloadable predictions

Project Structure

packages/
├── api/          # Hono API (Lambda)
│   └── src/
│       ├── services/   # Bedrock, SageMaker, DynamoDB, S3, code-gen
│       ├── routes/     # Upload, profiler, forecast, results, session
│       └── models/     # TypeScript types
├── web/          # React frontend
│   └── src/
│       ├── pages/      # Home, Upload, Analysis, Running, Results
│       └── hooks/      # SSE chat, session management
└── cdk/          # Infrastructure as Code
    └── lib/stacks/     # CloudFront, Lambda, DynamoDB, S3, WAF

Development

# Install dependencies
npm install

# Run API locally
cd packages/api && npm run dev

# Run frontend locally
cd packages/web && npm run dev

# Deploy
cd packages/cdk && npx cdk deploy

Security

• WAF WebACL with corp IP restriction (default BLOCK)
• API Gateway resource policy (CloudFront-only access)
• No direct API access — all traffic through CloudFront

License

MIT