Agentic SDLC Pipeline

AI-powered CI/CD pipeline that auto-reviews PRs with Claude, runs tests via UiPath Test Cloud, and orchestrates the full deploy flow through Maestro BPMN.

Comment

Agentic SDLC Pipeline 🚀

Hackathon: UiPath AgentHack 2026 — Track 2: UiPath Maestro BPMN

Live Webhook: uipath-agenthack.vercel.app/api/webhook

Repo: github.com/kyisaiah47/agentic-sdlc

Inspiration

Every engineering team has the same bottleneck: PRs sit waiting for a senior dev to review, tests run too late, and by the time a regression hits production the damage is done. The review-test-deploy cycle is slow, manual, and error-prone — not because engineers are bad at it, but because there's no system connecting all the pieces. UiPath Maestro BPMN is exactly that system. Paired with an AI code reviewer, it turns a chaotic handoff chain into a fully automated, auditable pipeline.

What It Does

Open a pull request — the pipeline takes it from there:

  • AI Code Review — a Claude Code agent reads the PR diff and returns a structured JSON verdict: approved or blocked, with specific line-level findings for security vulnerabilities, logic errors, and performance issues
  • Automated Testing — UiPath Test Cloud runs the full test suite against the PR branch in parallel
  • Smart Gateway — if both pass, the PR auto-merges to staging with no human needed; if either fails, the BPMN routes to a human approval gate
  • Human-in-the-Loop — a tech lead reviews Claude's findings and test failures inside Maestro, then approves or rejects
  • Production Deploy — once cleared, a GitHub Actions workflow fires via the dispatch API
  • Regression Monitoring — a post-deploy monitoring agent polls health endpoints; if consecutive failures are detected it triggers a rollback BPMN case automatically

How I Built It

Architecture

GitHub PR opened
      │  webhook
      ▼
Vercel Serverless Function (POST /api/webhook)
      │  verify signature, fetch diff
      ▼
UiPath Maestro BPMN Process (sdlc_pipeline.bpmn)
      │
      ├── [Step 1] Claude Code Agent — review PR diff
      │         returns { approved, blocking_issues, suggestions, summary }
      │
      ├── [Step 2] UiPath Test Cloud — run test suite on PR branch
      │         returns { passed, failed_tests, test_run_id }
      │
      ├── [Gateway] Both pass?
      │     ├── Yes → auto-merge to staging (GitHub API)
      │     └── No  → Human Review Gate (tech lead in Maestro)
      │                    ├── Approved → continue
      │                    └── Rejected → post GitHub comment, end
      │
      ├── [Step 3] GitHub Actions workflow_dispatch → production deploy
      │
      └── [Step 4] Monitoring Agent — polls health endpoints
                        └── regression detected → trigger RollbackBPMN case

Tech Stack

Layer Technology
AI Model Claude Sonnet 4.6 (Anthropic)
Code Review Agent Anthropic Python SDK, function calling
BPMN Orchestration UiPath Maestro BPMN
Automated Testing UiPath Test Cloud
Webhook Handler Python, Vercel Serverless
Deploy Trigger GitHub Actions workflow_dispatch API
Source Control GitHub (PR events, merge API)
Hosting Vercel

UiPath Components Used

Component Role
Maestro BPMN Orchestrates the full SDLC flow end-to-end
Agent Builder Claude Code reviewer agent, post-deploy monitoring agent
API Workflows GitHub webhook receiver, GitHub merge and deploy triggers
Test Cloud Automated test execution against each PR branch
UiPath for Coding Agents Used to build this solution and embedded in the review step

Challenges

  • Structured AI output: Getting Claude to return consistent JSON for every PR review — including edge cases like empty diffs or very large files — required careful prompt engineering and a markdown fence stripper for cases where the model wraps the response in code blocks.
  • BPMN gateway logic: Mapping the two-input gate (review result AND test result) cleanly in BPMN required an exclusive gateway with explicit condition expressions and a default flow, so the process never deadlocks when one agent is slow.
  • Webhook → BPMN handoff: GitHub delivers webhooks with a 10-second timeout. Fetching the PR diff and triggering Maestro both had to complete inside that window — solved by keeping the Vercel function lean and making the BPMN trigger async.

What's Next

  • Slack notifications at each BPMN gate so the tech lead gets pinged in-context
  • Multi-repo support — one Maestro process managing pipelines across an entire org
  • Historical review analytics: track which types of issues Claude catches most often per team
  • UiPath Test Cloud result drill-down surfaced directly in the Maestro human review task

Team

Built solo by @kyisaiah47 for UiPath AgentHack 2026, Track 2: Maestro BPMN.

Built With

  • anthropic-api
  • claude-sonnet-4.6
  • flask
  • github-actions
  • python
  • uipath-agent-builder
  • uipath-maestro-bpmn
  • uipath-test-cloud
  • vercel
Share this project:

Updates