{"slug": "show-hn-epochgo-local-first-multi-agent-automl-on-ollama-0-api-costs", "title": "Show HN: EpochGo – Local-first multi-agent AutoML on Ollama ($0 API costs)", "summary": "Epochlypse Research released EpochGo, an open-source local-first AutoML platform that runs three autonomous AI agents on Ollama with zero API keys and $0 cloud costs. The system uses CrewAI with the recommended qwen2.5-coder:7b model (~4.7 GB RAM) to generate Scikit-Learn preprocessing pipelines, benchmark 7 machine learning algorithms (XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, Gradient Boosting, and Logistic or Ridge Regression), and synthesize a FastAPI serving script, all via a single Docker Compose command on 16 GB RAM hardware. EpochGo targets developers who want private, offline data science automation without cloud dependencies.", "body_md": "**Autonomous, Local-First Multi-Agent AutoML Platform**\n\n*by [Epochlypse Research](https://epochlypse.com/) • Powered by CrewAI, Ollama, and Scikit-Learn / XGBoost / LightGBM / CatBoost*\n\n*Zero API keys. Zero cloud costs. 100% private and offline on your local machine.*\n\n**EpochGo** is an open-source, local-first AutoML software system developed by **Epochlypse Research**. Simply `git clone` the repository and run Docker. You get a localised link in your browser (`http://localhost:8000`), where you can create projects, drop in raw CSV datasets, and let a coordinated team of **three autonomous AI agents** do all the data science:\n\n1. \n🧹 **Agent 1: Data Preprocessor & Feature Architect**\n  - Automatically detects patterns, missing values, extreme outliers (IQR), categorical encodings, scaling, and high-dimensional multicollinearity (PCA).\n  - Generates and executes a standalone Scikit-Learn pipeline (`preprocessing.py` ).\n  - Produces a detailed before-vs-after comparison report (`preprocessing.md` ).\n  - Saves serialized transformer (`preprocessor.joblib` ) and cleaned dataset (`cleaned_data.csv` ).\n2. \n🏆 **Agent 2: Machine Learning Benchmark & Tuning Specialist**\n  - Evaluates **7 baseline machine learning algorithms** without hyperparameter bias:\n    - **Classification** : XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, Gradient Boosting, Logistic Regression.\n    - **Regression** : XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, Gradient Boosting, Ridge Regression.\n  - Evaluates baseline accuracy/F1/R²/RMSE and selects the **Top 3 contenders** .\n  - Runs high-efficiency hyperparameter optimization (`RandomizedSearchCV` ).\n  - Generates reproducible training script (`model_building.py` ) and comprehensive Markdown report (`model_building.md` ).\n  - Serializes the champion model to `model.joblib` .\n3. Evaluates \n4. \n🚀 **Agent 3: MLOps & Model Serving Engineer**\n  - Synthesizes a production-ready, standalone FastAPI microservice script (`serve.py` ) with Pydantic type validation matching your dataset schema.\n  - Enables an interactive, in-browser **Prediction Playground** to test predictions and probabilities in real time.\n  - Generates ready-to-use `curl` and Python API snippets.\n5. Synthesizes a production-ready, standalone FastAPI microservice script (\n\n- Docker & Docker Compose installed.\n- 16 GB RAM recommended (Apple Silicon M1/M2/M3/M4/M5 Mac, or Windows/Linux with 16GB RAM).\n\n```\ngit clone https://github.com/AryanDinakaran/EpochGo.git\ncd EpochGo\ndocker compose up --build\n```\n\nNavigate to:\n\n```\nhttp://localhost:8000\n```\n\nThat's it! EpochGo will automatically connect to Ollama, ensure the coding model is loaded, and start the web dashboard.\n\n| Metric | Target Specification | \n|---|---|\n| **Recommended Model** | **`qwen2.5-coder:7b`** | \n| **Why This Model?** | World-class code synthesis, scikit-learn mastery, and precise Markdown report writing. | \n| **VRAM / RAM Footprint** | ~4.7 GB RAM | \n| **Target Hardware** | 16 GB RAM Mac (Apple Silicon M-Series) or PC | \n| **Available Headroom** | Leaves >10 GB RAM free for parallel XGBoost / CatBoost / LightGBM model training | \n| **Low-Resource Fallback** | `qwen2.5-coder:1.5b` or`qwen2.5-coder:3b` for 8GB environments | \n\nTip\n\n**Already have Ollama installed on your Mac?**\nYou can save disk space and run native Metal GPU acceleration by pointing EpochGo to your existing Ollama instance:\n\n```\nOLLAMA_BASE_URL=http://host.docker.internal:11434 docker compose up\nphp\nflowchart TD\n    A[\"Raw CSV Upload & Project Name\"] --> B[\"Agent 1: Data Preprocessor\"]\n\n    subgraph Agent1 [\"🧹 Agent 1: Data Preprocessing\"]\n        B --> B1[\"Data Profiling & Pattern Analysis\"]\n        B1 --> B2[\"Imputation, Outlier Clipping, Encoding & Scaling\"]\n        B2 --> B3[\"Synthesize & Execute preprocessing.py\"]\n        B3 --> B4[\"Generate preprocessing.md & preprocessor.joblib\"]\n    end\n\n    B4 --> C[\"Agent 2: Model Builder\"]\n\n    subgraph Agent2 [\"🏆 Agent 2: Model Building & Tuning\"]\n        C --> C1[\"7 Baseline Algorithms Benchmark\"]\n        C1 --> C2[\"Select Top 3 Best Performers\"]\n        C2 --> C3[\"RandomizedSearchCV Hyperparameter Tuning\"]\n        C3 --> C4[\"Save Champion model.joblib\"]\n        C4 --> C5[\"Generate model_building.py & model_building.md\"]\n    end\n\n    C5 --> D[\"Agent 3: Model Serving\"]\n\n    subgraph Agent3 [\"🚀 Agent 3: Deployment & Serving\"]\n        D --> D1[\"Extract Pydantic Input Schema\"]\n        D1 --> D2[\"Synthesize Standalone serve.py (FastAPI)\"]\n        D2 --> D3[\"Activate In-Browser Inference Playground\"]\n    end\n```\n\nEvery project you create stores complete, reproducible artifacts inside `projects/<project_id>/`:\n\n| File | Purpose | \n|---|---|\n| `raw_data.csv` | Original uploaded dataset | \n| `cleaned_data.csv` | Transformed dataset after missing values, outliers, scaling & encoding | \n| `preprocessor.joblib` | Scikit-Learn transformer pipeline for downstream inference | \n| `preprocessing.py` | Standalone, reproducible Python script for data preparation | \n| `preprocessing.md` | Comprehensive Markdown report profiling before-and-after distributions | \n| `model.joblib` | Trained champion machine learning model artifact | \n| `model_building.py` | Standalone, reproducible model training and evaluation script | \n| `model_building.md` | Benchmark leaderboard (7 models), tuning stats, and feature importances | \n| `serve.py` | Production FastAPI microservice ready to deploy anywhere | \n| `metadata.json` | Project status, metric scoreboard, and feature schema | \n\nEpochGo includes two pre-configured datasets in `sample_data/` for immediate testing:\n\n1. \n**Customer Churn (`sample_data/customer_churn.csv`)** :\n  - **Task** : Binary Classification (`churn` : 0 or 1)\n  - Features: `age` ,`tenure_months` ,`monthly_charges` ,`total_charges` ,`contract` ,`internet_service` , etc.\n  - Includes real-world missing values and outliers to demonstrate agent cleaning.\n2. \n**House Prices (`sample_data/house_prices.csv`)** :\n  - **Task** : Regression (`price` )\n  - Features: `square_feet` ,`bedrooms` ,`bathrooms` ,`year_built` ,`neighborhood` ,`condition` , etc.\n\nIf you prefer to run EpochGo directly on your host machine:\n\n```\n# 1. Clone repository\ngit clone https://github.com/AryanDinakaran/EpochGo.git\ncd EpochGo\n\n# 2. Create virtual environment\npython3 -m venv venv\nsource venv/bin/activate\n\n# 3. Install dependencies\npip install -r requirements.txt\n\n# 4. Make sure Ollama is running\nollama serve\nollama pull qwen2.5-coder:7b\n\n# 5. Start EpochGo\nuvicorn epoch_go.main:app --host 0.0.0.0 --port 8000 --reload\n```\n\n- `GET /api/health` - Check service status and Ollama model connectivity.\n- `GET /api/projects` - List all projects and execution states.\n- `POST /api/projects` - Upload a CSV, name the project, and trigger the CrewAI pipeline.\n- `GET /api/projects/{id}` - Retrieve project metadata, status, and benchmark leaderboard.\n- `GET /api/projects/{id}/files/{filename}` - Retrieve generated Markdown reports and Python scripts.\n- `GET /api/projects/{id}/download/{filename}` - Download`.joblib` models or`.py` scripts.\n- `POST /api/projects/{id}/predict` - Run real-time predictions against the project's champion model.\n- `WS /ws/projects/{id}` - Real-time WebSocket stream for agent reasoning logs.\n\n- **Lead Architect** : Aryan Dinakaran\n- **Mission** : Local-first, private, reproducible multi-agent systems for high-leverage data science & machine learning engineering.\n- **Connect & Collaborate** :\n  - 🌐 Website: [epochlypse.com](https://epochlypse.com/)\n  - ✉️ Email: [founder@epochlypse.com](mailto:founder@epochlypse.com)\n  - 💼 LinkedIn: [Aryan Dinakaran](https://www.linkedin.com/in/aryan-dinakaran-462277230/)\n  - 📸 Instagram: [@aryandinakaran](https://www.instagram.com/aryandinakaran/)\n  - 🐙 GitHub: [@AryanDinakaran](https://github.com/AryanDinakaran)\n- 🌐 Website: \n\nMIT License. Free for personal, commercial, and research use.", "url": "https://wpnews.pro/news/show-hn-epochgo-local-first-multi-agent-automl-on-ollama-0-api-costs", "canonical_source": "https://github.com/AryanDinakaran/EpochGo", "published_at": "2026-10-12 10:20:08+00:00", "updated_at": "2026-10-12 10:39:42.619295+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "machine-learning", "developer-tools", "ai-products"], "entities": ["EpochGo", "Epochlypse Research", "Ollama", "CrewAI", "qwen2.5-coder:7b", "XGBoost", "LightGBM", "CatBoost"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/show-hn-epochgo-local-first-multi-agent-automl-on-ollama-0-api-costs", "markdown": "https://wpnews.pro/news/show-hn-epochgo-local-first-multi-agent-automl-on-ollama-0-api-costs.md", "text": "https://wpnews.pro/news/show-hn-epochgo-local-first-multi-agent-automl-on-ollama-0-api-costs.txt", "jsonld": "https://wpnews.pro/news/show-hn-epochgo-local-first-multi-agent-automl-on-ollama-0-api-costs.jsonld"}}