AI Lifecycle

A structured agentic team for the full AI/ML project lifecycle — a BMad Method module.

Five specialist agents guide you from raw domain research through production inference. Works across paradigms — deep learning, gradient boosting (XGBoost/LightGBM), transformers, fine-tuning, classical ML, and hybrid approaches.

Meet the Team

Agent	Name	Role
`ai-agent-domain-expert`	Alex	Researches the problem domain, frames the Research Thesis, writes the PRD, and audits upstream docs after each experiment cycle.
`ai-agent-data-engineer`	Sam	Performs EDA, characterizes data quality and distribution, and establishes a model-agnostic performance baseline.
`ai-agent-researcher`	Maya	Selects the modelling paradigm and architecture, designs the experiment strategy, evaluates results against TECHSPEC tiers, and captures learnings.
`ai-agent-mlops-engineer`	Kai	Writes TECHSPEC contracts, builds data and training pipelines, and adapts trained models for production deployment.
`ai-agent-experimentation-engineer`	Jordan	Executes training and fit runs, logs everything to the experiment tracker, and runs hyperparameter optimization.

Lifecycle at a Glance

[0]   Setup        ai-setup configure / new-project
[1]   Research     Alex   — domain-research     → Domain Knowledge Base
[1.5] Ideation     Alex   — ideation            → Research Thesis + PRD
[2]   EDA          Sam    — eda                 → EDA Report + baseline
[3]   Architecture Maya   — architecture        → Architecture document
[4]   Design       Maya   — detailed-design     → INF-* + EXP-* tasks
[4.5] TECHSPEC     Kai    — techspec            → Pre-experiment contract
[5]   Infra        Kai    — infra               → Pipelines + eval harness
[6]   Experiment   Jordan — experiment          → Experiment Log + configs
[6.5] Results      Jordan — results            → Raw metrics, curves, comparisons
[7]   Analysis     Maya   — analysis            → Interpretation + lessons + next steps
      ├── [7.5]    Jordan — hparam (if needed)  → HPO report → back to [6]
      └── [8]      Alex   — revision-audit      → Revision Log → back to [4.5]

Anytime:
      Alex   — advise             Surface validated params from past experiments
      Kai    — decisions          Capture know-how and rejected alternatives
      Kai    — inference-pipeline Adapt model for production + V&V

Getting Started

Prerequisites

Requirement	Notes
BMad Method	Required — provides `/bmad-help` routing and `_bmad/` structure
AI IDE	Claude Code, Antigravity, or VSCode + Cline / Cursor
Claude Sonnet 4.6+ or equivalent	Recommended model for all agents
uv	Python package manager — `curl -LsSf https://astral.sh/uv/install.sh \| sh`

Install the module

Use the BMad installer from your project root. Choose the method that matches your environment:

From a Git repository or npm registry (default / Artifactory)

npx bmad-method install \
  --directory . \
  --modules bmm \
  --custom-source https://github.com/avielbl/ai-lifecycle \
  --tools claude-code \
  --yes

If your organisation mirrors npm packages through an internal Artifactory or similar registry, replace the GitHub URL with the package name:

npx bmad-method install \
  --directory . \
  --modules bmm \
  --custom-source ai-lifecycle \
  --tools claude-code \
  --yes

From a local path (offline / air-gapped environments)

npx bmad-method install \
  --directory . \
  --modules bmm \
  --custom-source /path/to/ai-lifecycle \
  --tools claude-code \
  --yes

Interactively

npx bmad-method install
# When prompted for a custom source → paste the GitHub URL or local path

Configure (once per project)

Run the setup skill to write config and register capabilities with /bmad-help:

/ai-setup   → then: configure

Or headless with defaults:

/ai-setup configure --headless

Scaffold a new project (optional)

For brand-new projects, scaffold the full directory structure, IDE config, and uv project:

/ai-setup   → then: new-project

This creates data/, src/, notebooks/, configs/, docs/, models/, outputs/, a pyproject.toml (no dependencies yet), and IDE-specific agent config.

No packages are installed at scaffold time. Dependencies are added in Ideation (Stage 1.5) and installed in Infrastructure (Stage 5) via uv sync.

Navigate with BMad Help

/bmad-help

Reads your project state and tells you exactly which agent and capability to invoke next.

Invoking Agents

Claude Code

The install step copies agents into .claude/skills/, enabling slash commands:

/ai-agent-domain-expert   → tell it which capability to activate
/ai-agent-data-engineer
/ai-agent-researcher
/ai-agent-mlops-engineer
/ai-agent-experimentation-engineer

Antigravity

Auto-discovered globally. Use slash commands directly.

VSCode + Cline / Cursor

No slash commands. Reference the agent by path:

Follow the workflow in: .claude/skills/ai-lifecycle/ai-agent-domain-expert/SKILL.md
Activate capability: domain-research

The .clinerules file generated by ai-setup new-project lists all agent paths for quick copy-paste.

Stages Reference

Sequential Lifecycle

Stage	Agent	Capability	Output
1	Alex (domain-expert)	`domain-research`	Domain Knowledge Base
1.5	Alex (domain-expert)	`ideation`	Research Thesis + PRD
2	Sam (data-engineer)	`eda`	EDA Report
3	Maya (researcher)	`architecture`	Architecture document
4	Maya (researcher)	`detailed-design`	Detailed Design (INF-* + EXP-* tasks)
4.5	Kai (mlops-engineer)	`techspec`	TECHSPEC contract
5	Kai (mlops-engineer)	`infra`	Pipelines + eval harness
6	Jordan (experimentation-engineer)	`experiment`	Experiment Log + archived configs
6.5	Jordan (experimentation-engineer)	`results`	Raw metrics, curves, comparison tables
7	Maya (researcher)	`analysis`	Interpretation + lessons + next steps
7.5 (conditional)	Jordan (experimentation-engineer)	`hparam`	HPO report
8	Alex (domain-expert)	`revision-audit`	Revision Log + amended upstream docs

Anytime Capabilities

Agent	Capability	When to Use
Alex (domain-expert)	`advise`	Before any experiment — surfaces validated params and dead ends from past work
Kai (mlops-engineer)	`decisions`	During any experiment — captures know-how, rejected alternatives, and deferred questions
Kai (mlops-engineer)	`inference-pipeline`	After a model is accepted — adapts for production with V&V

Task Namespaces

Detailed Design (Stage 4) produces two task categories:

Prefix	Executed By	Purpose
`INF-*`	Kai — `infra`	Infrastructure tasks — built once, reused across experiment cycles
`EXP-*`	Jordan — `experiment`	Experiment tasks — executed each cycle against the active TECHSPEC
`REV-*`	Next cycle	Generated by `revision-audit` for the following iteration

Experiment Tracking

Tool	Best For
Weights & Biases	Teams, sweep UI, collaboration
MLflow	Self-hosted, open-source, model registry
ClearML	Auto-capture, enterprise MLOps, HPO orchestration

Choose one during ai-setup configure; wire it in infra (Stage 5).

LLM Configuration

Agent model

Set in your IDE — not here. Any model supported by your IDE works.

Script-level LLM calls

Some utility scripts call an LLM directly via scripts/llm_client.py, configured by configs/llm_config.yaml (written by ai-setup new-project):

# Anthropic / Claude (default)
provider: anthropic
model: claude-sonnet-4-6
base_url: ~
api_key_env: ANTHROPIC_API_KEY

# OpenAI-compatible
# provider: openai-compatible
# model: gpt-4o
# base_url: http://localhost:11434/v1
# api_key_env: OPENAI_API_KEY

The API key lives in the env var — never in the config file.

Key Principles

Configure first. Run ai-setup configure before invoking any agent.
Scaffold, then install. No uv sync until Kai runs infra (Stage 5).
TECHSPECs are contracts. Lock them before training starts; amend them with a new revision.
HPO only after baseline confirmation. HPO on a broken architecture wastes compute.
Run advise before every experiment. Alex mines past projects so you don't repeat mistakes.
Document decisions as you go. Kai captures rejected alternatives and know-how in DECISIONS.md — don't lose this context.
Fresh context window per agent. Each agent is specialized — mixing stages in one session degrades quality.
Failed attempts are mandatory. Infra Log and Analysis documents with no failures documented are incomplete.

CI/CD

Two GitHub Actions workflows are included:

validate_skills.yml — PR gate: checks SKILL.md frontmatter, manifest structure, and semver format
update_marketplace.yml — auto-generates marketplace.json from all manifests on push to main

Updating the Module

Re-run the installer with the same --custom-source you used during installation. BMad re-fetches from the original source and applies updates in place.

npx bmad-method install \
  --directory . \
  --modules bmm \
  --custom-source https://github.com/avielbl/ai-lifecycle \
  --tools claude-code \
  --yes

For offline environments, use the local path instead of the GitHub URL.

Versioning

v4.1.0 — Renamed MLOps Developer and Experimentation Engineer personas for role clarity. Restructured experiment output into per-experiment folders with dedicated results and decisions capabilities; merged retrospective into analysis and dropped numeric prefixes. Added package.json for npm/registry distribution with environment-specific install examples (GitHub, Artifactory, local path). Repaired broken file references across skills. Updated docs to use BMad installer instead of git submodules.
v4.0.0 — Renamed from bmad-dl-lifecycle to ai-lifecycle. Module code ai. Broadened from deep learning to all AI/ML paradigms. Agents renamed to ai-agent-* with assigned personas (Alex, Sam, Maya, Kai, Jordan). ai-setup skill absorbs scaffold and module configuration.
v3.0.0 — Agent-based architecture. Five domain specialists replace per-skill approach. Memory added to Domain Expert, Researcher, Developer.
v2.1.0 — Scaffold (Stage 0): automated project scaffolding with uv.
v2.0.0 — Split infra + experiment. HPO (Stage 7.5). W&B/MLflow/ClearML integration.
v1.2.0 — Knowledge flywheel: advise, techspec.
v1.0.0 — Initial release.

License

MIT

Navigation Menu

Repository files navigation

AI Lifecycle

Meet the Team

Lifecycle at a Glance

Getting Started

Prerequisites

Install the module

Configure (once per project)

Scaffold a new project (optional)

Navigate with BMad Help

Invoking Agents

Claude Code

Antigravity

VSCode + Cline / Cursor

Stages Reference

Sequential Lifecycle

Anytime Capabilities

Task Namespaces

Experiment Tracking

LLM Configuration

Agent model

Script-level LLM calls

Key Principles

CI/CD

Updating the Module

Versioning

License

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 45 Commits
.claude-plugin		.claude-plugin
.github/workflows		.github/workflows
ai-agent-data-engineer		ai-agent-data-engineer
ai-agent-domain-expert		ai-agent-domain-expert
ai-agent-experimentation-engineer		ai-agent-experimentation-engineer
ai-agent-mlops-engineer		ai-agent-mlops-engineer
ai-agent-researcher		ai-agent-researcher
ai-setup		ai-setup
scripts		scripts
README.md		README.md
ai-lifecycle.csv		ai-lifecycle.csv
module.yaml		module.yaml
package.json		package.json

Folders and files

Latest commit

History

Repository files navigation

AI Lifecycle

Meet the Team

Lifecycle at a Glance

Getting Started

Prerequisites

Install the module

Configure (once per project)

Scaffold a new project (optional)

Navigate with BMad Help

Invoking Agents

Claude Code

Antigravity

VSCode + Cline / Cursor

Stages Reference

Sequential Lifecycle

Anytime Capabilities

Task Namespaces

Experiment Tracking

LLM Configuration

Agent model

Script-level LLM calls

Key Principles

CI/CD

Updating the Module

Versioning

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages