Why Gaslamp?

The landscape of AI coding agents has evolved rapidly. Tools like Cursor and Claude Code are incredibly powerful at generating code. However, without proper guiderails, these agents often hallucinate architectures, reinvent the wheel, or use outdated practices.

Gaslamp exists to solve the "Blank Canvas Problem."

When you ask an agent to "build a financial data pipeline," it has a million ways to do it. Gaslamp provides the agent with Professional Machine Learning Blueprints.

Instead of guessing, the agent builds exactly what you need, using proven architectures, predefined schemas, and rigorous evaluation guardrails.

Core Values

1. Guardrails over Suggestions: We don't just provide "ideas." Gaslamp blueprints come with strict rules of engagement (data_gotchas, schema requirements) that the LLM must obey.
2. Vendor Independence: We build architectures that run anywhere. We avoid lock-in to specific proprietary cloud platforms unless absolutely necessary.
3. Domain Expertise: Our blueprints encode years of professional data science and ML engineering experience right into the context window.

Installation & Setup

Gaslamp is designed to be lightweight and zero-friction. You do not need to manually initialize project spaces or configure complex MCP server connections upfront. Gaslamp handles everything on demand.

Prerequisites

You need a capable AI coding agent installed. We recommend:

• Claude Code (CLI-based)
• Cursor (IDE-based)

1. Install the Skill

Simply install the Gaslamp master skill into your agent's environment. (Note: specific install commands depend on your platform's skill registry).

2. Start Building

Everything happens on-demand. Just open your agent and invoke the orchestrator:

/gaslamp I want to build a churn prediction model

The agent will automatically create the necessary project workspace. If a specific phase of your project requires an MCP server (like pulling a domain blueprint or generating a UI), the agent will pause and provide you with the exact 1-line command to install it right when you need it.

The Master Orchestrator

Gaslamp is not a generic AI agent. It acts as an Orchestrator Master Skill for your existing coding agent (like Claude Code or Cursor).

Initialize a project by typing:

/gaslamp start a new churn prediction project

When invoked, the /gaslamp orchestrator analyzes your request and automatically delegates the work to one of five highly-specialized AI Sub-Skills, moving your project sequentially through the ML lifecycle.

The 5 Sub-Skills

Gaslamp ships with 5 encoded sub-skills. Your agent assumes these personas depending on the current phase of your project.

1. ml-buddy (Ideation to Evaluation)

The core data science pair programmer. It handles the initial interview, formulates a data strategy, selects a model architecture ("No-BS Model Choice"), and runs rigorous evaluations.

2. deploy-buddy (Optimization to Serving)

Steps in when your model is trained. It selects the best inference engine (ONNX, TensorRT, vLLM) and auto-generates serving code (FastAPI, Triton).

3. report-buddy (Reporting & Demonstration)

The PM-focused companion. It reads your project logs to synthesize business-ready Executive Summaries and statically compiled Interactive HTML Demos for stakeholders.

4. unsloth-buddy (LLM Fine-Tuning)

Invoked specifically for Large Language Model tasks. It bypasses standard training loops to utilize the highly optimized Unsloth library, ensuring 4-bit quantization and PEFT/LoRA patching to prevent OOM errors.

5. pipeline-buddy (Cloud Pipelines)

The data engineering companion. It translates your local, monolithic ML scripts into production-ready cloud pipelines (Tangle) by generating component .yaml files and connectivity_check.py scripts.

Standardized State Management

Instead of relying on an agent's massive (and easily confused) context window, Gaslamp forces the agent to maintain State locally on disk:

• gaslamp.md: The global project state. This file lives at the root of your project directory and records major architectural decisions, selected models, and overall progress. All buddies read this file.
• [skill]-buddy/progress_log.md: An append-only log of granular actions taken by a specific sub-skill during a session.
• [skill]-buddy/memory.md: Technical scratchpad for hyperparameters, dataset shapes, and debugging discoveries.

MCP Dependencies

Gaslamp deeply integrates with the Model Context Protocol (MCP) to provide agents with external context and actions. Different sub-skills require different MCP servers to function:

gaslamp_tech_pack_fetcher

• Used By: ml-buddy
• Purpose: Allows the agent to fetch "Tech Packs" (Domain Blueprints) directly from the Gaslamp registry. For example, if you are building an Options Pricing Engine, ml-buddy will use this tool to pull the finance blueprint (containing recipes and data_gotchas) before writing code.
• Install Command: npx -y @gaslamp/tech-pack-fetcher

gaslamp_demo_builder

• Used By: report-buddy
• Purpose: Allows the agent to fetch interactive HTML/CSS UI templates (e.g., a "dark-mode" classification dashboard) and inject your model's outputs into a standalone browser demo.
• Install Command: npx -y gaslamp-demo-builder

Note: The Gaslamp orchestrator uses a "lazy failure" pattern. If an agent tries to invoke a tool and fails, it will pause, instruct you to run the specific npx install command to register the MCP server, and then wait for you to restart the session.