AI Builder

From an AI idea to a working, evaluated system.

The AI Builder helps you choose an implementation path based on your skills, requirements, available models, information, and deployment environment.

If you haven't decided what kind of AI solution is appropriate, begin with AI Solver.

Choose Components for Your Implementation

What component of an AI system must I choose, connect, or operate?

Start with AI Tools & Infrastructure Navigator, or choose a specific branch:

• Choose libraries and programming tools
• Connect applications and data services
• Adapt or compress a model
• Plan deployment and lifecycle tooling
• Build a reproducible prototype

Start: What Are You Trying to Build?

Your Goal	Follow This Branch
I want a quick experiment or proof of concept.	A. Prototype
I want to build an AI-powered application.	B. Application Development
I want to customize an AI model.	C. Model Customization
I want a system that uses tools and performs tasks.	D. Agents and Automation
I need to connect AI with my data and software.	E. Integration and Architecture
I need to deploy and operate a system.	F. Deployment and Operations

A. Prototype

Question: Can the initial idea be tested without custom programming?

• Yes: Google AI Studio → Codeless Options, Code Generators, Drag n' Drop
• I want to experiment with models: Hugging Face → Model Zoos
• I want a visual machine-learning experiment: TensorFlow Playground → Teachable Machine
• I want to build an experimental workflow: Development → Notebooks
• I need programmatic control: Application Development →

Decision: Has the prototype demonstrated useful results on representative examples?

• No: Revisit the problem definition in AI Solver, improve the inputs, or use a simpler baseline.
• Yes: Write down the required inputs, outputs, constraints, and evaluation criteria using Requirements Management.

Checkpoint: Demonstrate the system on multiple representative inputs, including difficult cases.

B. Application Development

Question: What development approach fits your requirements?

• Python and machine-learning workflows: Python → Generative AI with Python → PyTorch
• Web and browser-based applications: JavaScript → Generative AI with JavaScript → TensorFlow.js
• Deep-learning model development: TensorFlow → PyTorch → Libraries & Frameworks
• Interactive notebooks: Notebooks → Jupyter
• AI-assisted programming: AI Pair Programming Tools
• Existing reusable components: Libraries & Frameworks Overview

Decision: Does your application need to learn its own model parameters?

• No: Start by integrating an existing model or service.
• Yes: Evaluate model customization →

Checkpoint: Can your application handle valid inputs, incorrect inputs, service failures, and repeatable tests?

Continue: Integration and Architecture →

C. Model Customization

Question: Why isn't the existing model sufficient?

• It needs clearer instructions: Prompt Engineering (PE) → Context
• It needs access to proprietary information: Retrieval-Augmented Generation (RAG) → Database
• It needs to use external tools: Agents/Assistants → LangChain
• It needs consistent specialist behavior: Fine-tuning → Train, Validate, and Test
• It needs lower runtime cost or a smaller footprint: Distillation → Quantization
• It needs an entirely different architecture: Neural Architecture → Algorithms

Decision: Can prompting, retrieval, or tool integration solve the problem without additional training?

• Yes: Prefer the simpler approach and evaluate it.
• No: Determine whether suitable training data and resources are available before fine-tuning or training.

Checkpoint: Show that customization improves an agreed evaluation metric over the original model.

D. Agents and Automation

Question: How much decision-making does the system require?

• A repeatable sequence of fixed steps: Robotic Process Automation (RPA) → Web Automation
• AI processing within a fixed workflow: Development → LangChain
• Dynamic tool selection and multi-step execution: Agents/Assistants → LangChain
• Actions involving physical environments: Robotics → Embodied AI
• Repeated decisions using feedback: Reinforcement Learning (RL) → Policy

Decision: Can a controlled workflow achieve the same result with less complexity?

• Yes: Use the workflow.
• No: Define the agent's tools, permissions, memory, stopping conditions, and human review requirements.

Checkpoint: Can the system recover from tool errors without producing uncontrolled or unauthorized actions?

Continue: Dimension 7: Trust & Evaluation →

E. Integration and Architecture

Question: What must the AI system connect to?

• Documents or knowledge repositories: Retrieval-Augmented Generation (RAG) → LlamaIndex
• Databases: Database → Data Interoperability
• Applications, APIs, and services: Development → Service Capabilities
• Cloud infrastructure: Platforms: AI/Machine Learning as a Service (AIaaS/MLaaS)
• Devices and sensors: Internet of Things (IoT) → Digital Twin
• An organization-wide architecture: Enterprise Architecture (EA) → Generative AI Stack

Decision: Where must information and computation reside?

• Locally or on a device: Examine hardware, model size, and deployment constraints.
• In a cloud service: Evaluate privacy, availability, cost, and integration.
• Across multiple environments: Evaluate interoperability, monitoring, and failure handling.

Continue: Dimension 6: Technology Selection →

F. Deployment and Operations

Question: What operating environment is required?

• Experimental environment: Notebooks → Building Your Environment
• Cloud-hosted application: Platforms: AI/Machine Learning as a Service (AIaaS/MLaaS)
• Managed model lifecycle: AIOps/MLOps
• Continuous monitoring: Model Monitoring
• Containers and distributed services: Containers; Docker, Kubernetes & Microservices
• Small or on-device model: Quantization → TensorFlow Lite

Decision: Is the system operationally ready?

Verify:

1. Representative test cases pass.
2. Sensitive information is adequately protected.
3. Costs and performance meet requirements.
4. Errors are logged and monitored appropriately.
5. A rollback or shutdown procedure exists.
6. There is an accountable human owner.

Continue: Evaluate the system →

Builder Completion Check

A build isn't complete merely because the model produces an answer.

The system should have defined requirements, a working implementation, appropriate test coverage, operational safeguards, and a maintenance plan.

Related decision trees:

• AI Solver — identify the solution approach.
• Data & Knowledge Navigator — prepare its information.
• AI Technology Selector — choose products and platforms.
• AI Trust & Evaluation — establish acceptance criteria.
• AI Strategy & Adoption — manage the project and investment.

Home · ConceptChains · Primo.ai Pages