Autonomous AI-Driven SDLC Orchestration — From PRD to Executive Dashboard, fully orchestrated by Claude-powered agents.
FlowForge AI is a conceptual agentic platform that demonstrates how an AI-native engineering team could automate the entire Software Development Lifecycle (SDLC) — from parsing a Product Requirements Document all the way to delivering a voice-narrated executive briefing — using specialized Claude-powered agents at every stage.
This project was designed to showcase:
- Systems-level AI architecture applied to a real-world engineering problem
- Agentic orchestration patterns using the Anthropic Claude API
- End-to-end SDLC automation across 5 distinct workflow layers
┌──────────────────────────────────────────────────────────────────────────┐
│ FlowForge AI — 5 Layers │
├──────────────┬──────────────┬──────────────┬─────────────┬───────────────┤
│ LAYER 01 │ LAYER 02 │ LAYER 03 │ LAYER 04 │ LAYER 05 │
│ Ingestion │ Planning │ Execution │ Review │ Reporting │
├──────────────┼──────────────┼──────────────┼─────────────┼───────────────┤
│ PRD Upload │ FRD → Sprint │ Ticket → │ PR Diff → │ Metrics → │
│ ↓ │ Backlog │ Code │ Review │ Dashboard │
│ FRD + ERD │ ↓ │ ↓ │ ↓ │ ↓ │
│ Generated │ Jira Tickets │ PR + Defects │ Gate Check │ Voice Brief │
└──────────────┴──────────────┴──────────────┴─────────────┴───────────────┘
All layers powered by Anthropic Claude API
| Agent | Input | Output |
|---|---|---|
| PRD Ingestion Agent | Raw PRD (PDF/DOCX/MD) | Normalized JSON schema |
| FRD Generation Agent | Requirements JSON | Formatted FRD with acceptance criteria |
| ERD Synthesis Agent | FRD content | Mermaid ERD diagram |
| Agent | Input | Output |
|---|---|---|
| Sprint Orchestrator Agent | Approved FRD | Fibonacci-estimated sprint backlog |
| Jira Writer Agent | Sprint backlog | Jira epics, stories, subtasks |
| Dependency Mapper Agent | Jira backlog | Blocking/blocked-by relationships |
| Agent | Input | Output |
|---|---|---|
| Code Generation Agent | Jira ticket + ERD | Scaffolded code + unit test stubs |
| PR Creation Agent | Feature branch | GitHub PR with full metadata |
| Defect Detection Agent | CI/CD webhook | Jira bug ticket with failure log |
| Assignment Agent | Ticket + team profiles | Optimal developer assignment |
| Agent | Input | Output |
|---|---|---|
| Diff Analyzer Agent | PR diff | Structured diff analysis JSON |
| Code Review Agent | Diff + ACs | Inline PR comments + summary |
| Approval Gate Agent | Quality score | Ready / Needs Work status |
| Agent | Input | Output |
|---|---|---|
| Summary Aggregation Agent | Jira + GitHub + CI/CD | Weekly metrics JSON |
| Dashboard Rendering Agent | Metrics JSON | Interactive HTML dashboard |
| Voice Briefing Agent | Metrics summary | 10-min MP3 briefing + transcript |
| Q&A Routing Agent | Exec question | Slack/email to PM or assignee |
Full PRDs for all 5 layers are included in this repository:
| PRD | Layer | Contents |
|---|---|---|
| PRD-01 | Requirements Ingestion | Goals, agent architecture, functional & non-functional requirements |
| PRD-02 | Sprint Planning & Jira | Orchestration logic, Jira API integration specs |
| PRD-03 | Code Generation & Assignment | Coding agents, PR creation, defect detection |
| PRD-04 | PR Diff Review | Quality gate logic, review agent design |
| PRD-05 | Executive Dashboard | Voice briefing, Q&A routing, exec UX |
All agents communicate through the Anthropic /v1/messages endpoint:
const response = await fetch("https://api.anthropic.com/v1/messages", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({
model: "claude-sonnet-4-20250514",
max_tokens: 1000,
system: AGENT_SYSTEM_PROMPT, // Agent-specific instructions
messages: [
{ role: "user", content: agentInput }
]
})
});Each agent has a dedicated system prompt that defines its role, input contract, output format, and reasoning constraints. System prompts are version-controlled and configurable per team.
| Metric | Before FlowForge AI | After FlowForge AI |
|---|---|---|
| PRD → Jira backlog time | 2–3 days | < 10 minutes |
| Sprint planning ceremony | 3–4 hours | < 45 minutes |
| First PR review turnaround | 8+ hours | < 3 minutes |
| Pre-merge defect catch rate | Baseline | +30% improvement |
| Weekly exec status prep | 2–4 hours engineering time | Fully automated |
| Executive briefing duration | 30-minute meeting | 10-minute voice brief |
| Component | Technology |
|---|---|
| AI Agent Engine | Anthropic Claude API (claude-sonnet-4-20250514) |
| Orchestration | LangGraph-style agent graph (node-per-agent) |
| Sprint Management | Jira Cloud REST API v3 |
| Source Control | GitHub / GitLab (PR hooks + webhooks) |
| CI/CD Integration | GitHub Actions / Jenkins webhook events |
| ERD Rendering | Mermaid.js |
| Voice Synthesis | ElevenLabs API |
| Dashboard | React + Recharts, published to exec portal |
| Notifications | Slack API + Email |
FlowForge AI is designed with human oversight at every critical gate:
- FRD Review Gate — PM approves FRD before sprint planning begins
- Sprint Commitment Gate — Team confirms sprint backlog before Jira tickets are created
- Code Review Gate — Agent provides first-pass review; human engineer must approve merge
- PR Merge Gate — No automated merges; all merges require human sign-off
- Q&A Routing — Exec questions routed to humans, not answered by AI directly
AI accelerates. Humans decide.
This project applies key concepts from MIT Applied AI coursework:
- Specialization over generalization — Each agent has a single, well-defined responsibility
- Structured I/O contracts — Agents communicate via JSON schemas, not freeform text
- Confidence-aware output — Low-confidence generations are flagged for human review
- Feedback loops — Dismissed agent comments and gate overrides are logged as training signal
- Graceful degradation — If an agent fails, the pipeline surfaces the error rather than silently skipping
flowforge-ai/
├── README.md
├── docs/
│ ├── PRD-01-Requirements-Ingestion.md
│ ├── PRD-02-Sprint-Planning.md
│ ├── PRD-03-Code-Generation.md
│ ├── PRD-04-PR-Review.md
│ └── PRD-05-Exec-Dashboard.md
├── agents/
│ ├── layer01-ingestion/
│ │ ├── prd-ingestion-agent.js
│ │ ├── frd-generation-agent.js
│ │ └── erd-synthesis-agent.js
│ ├── layer02-planning/
│ ├── layer03-execution/
│ ├── layer04-review/
│ └── layer05-reporting/
├── prompts/
│ └── system-prompts/ # Agent system prompt templates
├── schemas/
│ └── io-contracts/ # JSON I/O schemas per agent
└── demo/
└── mock-data/ # Sample PRDs, FRDs, sprint data
# Clone the repo
git clone https://github.com/yourusername/flowforge-ai.git
cd flowforge-ai
# Install dependencies
npm install
# Set your Anthropic API key
export ANTHROPIC_API_KEY=your_key_here
# Run the demo pipeline with sample PRD
node demo/run-pipeline.js --input demo/mock-data/sample-prd.mdThis project was published as a LinkedIn carousel post demonstrating AI-native SDLC orchestration. The carousel and full architecture diagram are available in the /assets folder.
Built to demonstrate systems-level AI architecture and agentic workflow design, grounded in MIT Applied AI coursework and real-world engineering team experience.
Connect on LinkedIn | Star this repo | Drop a comment on the post
MIT License — feel free to fork, extend, and build on this concept.
FlowForge AI is a conceptual demo. It is not a production system. All projected metrics are estimates based on industry benchmarks and team experience.