🚀 Building an AI-Powered Code Reviewer for Bitbucket Using Groq & Pipelines

Modern development teams rely heavily on pull requests for code quality-but manual reviews are slow, inconsistent, and expensive. Recently, Bitbucket introduced Rovo Dev, and GitHub has Ask Copilot, both offering AI-assisted PR reviews.

But there was one major problem for me:

❌ I wasn’t ready to pay $20 per developer per month just to get AI reviews.

✅ I already had a Groq API key.

✅ I wanted a fully automated, pipeline-driven solution.

So I built my own AI-powered PR review system for Bitbucket using:

• ✅ Bitbucket Pipelines
• ✅ Groq LLM (llama-3.3-70b-versatile)
• ✅ Git-based diff extraction (no REST API auth headaches)

This system reviews every PR automatically and outputs a structured, checklist-driven AI review-with zero dependency on Bitbucket’s unreliable token ecosystem and zero per-developer licensing cost.

In this post, I’ll cover:

• How this compares to Rovo Dev & GitHub Copilot
• Why I avoided Bitbucket’s REST APIs
• The final production architecture
• How the AI review works
• Key engineering lessons from building this

---

🤖 Rovo Dev vs Ask Copilot vs This Approach

Feature	Rovo Dev (Bitbucket)	Ask Copilot (GitHub)	This Groq-Based System
AI PR Reviews	✅	✅	✅
Fully Automated in CI	❌ (mostly UI based)	❌ (manual prompts)	✅
Per-Developer Cost	❌ $20/month/dev	❌ Bundled with Copilot	✅ $0 per dev
Works in Pipelines	❌	❌	✅
Custom Review Rules	❌ Limited	❌ Limited	✅ Full control
Vendor Lock-in	✅	✅	❌ None (Groq + Git)

I didn’t want:

• Another per-seat SaaS subscription
• A manual “Ask AI” workflow
• Or a system that breaks when pricing changes

I wanted:
✅ Fully automated

✅ CI-level enforcement

✅ Custom review rules

✅ Lowest possible cost

That’s why I chose Groq + Pipelines.

---

❌ The Problem with Traditional Bitbucket PR Automation

Initially, I tried the standard approach:

• Fetch PR diffs using the Bitbucket REST API
• Post PR comments using:
  • Atlassian API tokens
  • Workspace tokens
  • Repository access tokens

Despite correct scopes, PR comment posting repeatedly failed with 401 Unauthorized errors due to:

• Inconsistent token behaviors
• Bitbucket’s evolving security model
• Poor documentation around 2025 token behavior

After continous debugging, I realized:

✅ The smartest move was to eliminate Bitbucket’s REST API entirely for diff collection.

---

✅ The Final Working Architecture

Here’s the production setup that actually works:

Pull Request Created
↓
Bitbucket Pipeline Triggered
↓
Git diff extracted using: git diff origin/main...HEAD
↓
Diff sent to Groq LLM
↓
AI generates structured checklist-based review
↓
Review shown in Pipeline logs + downloadable artifact

Why this works so well:

• ✅ No REST API calls for diffs
• ✅ No authentication failures
• ✅ No permission issues
• ✅ No flakiness
• ✅ Fully deterministic

---

🤖 The AI Review Rules (Enterprise-Grade)

The AI review is driven by a strict TypeScript + Angular + security checklist:

• ❌ No any types
• ✅ Strong typing with interfaces & generics
• ✅ Modern Angular syntax (@if, @for, standalone components)
• ✅ Authentication guards
• ✅ No hardcoded secrets
• ✅ Error handling
• ✅ Tests present
• ✅ Performance checks
• ✅ Accessibility (WCAG)
• ✅ Final verdict: MERGE READY / NEEDS WORK

This ensures:

• Consistent reviews
• Enforced standards
• Zero reviewer bias

---

🧠 Git-Based Diff Instead of REST API

Instead of calling Bitbucket’s REST endpoints, the pipeline simply runs:

git fetch origin main
git diff origin/main...HEAD

This gives:

✅ The exact PR diff

✅ No API authentication

✅ Works in every CI environment

This single decision eliminated 90% of the system’s complexity.

⚡ Groq LLM Integration
The diff is sent to Groq using:
llama-3.3-70b-versatile

Why Groq?
⚡ Extremely fast inference

🧠 Excellent reasoning on large diffs

💸 Much cheaper than many alternatives

✅ OpenAI-compatible API

🌱 More eco-friendly due to lower compute time per request

The AI responds with:

🚨 Critical Issues

🔒 Security Analysis

⚡ Performance Review

🏗️ Architecture Feedback

📝 Maintainability

✅ Final Verdict: MERGE READY / NEEDS WORK

📄 Where the AI Review Appears Instead of battling PR comment permissions:

✅ The full AI review appears in the Pipelines logs

✅ Optionally saved as a downloadable ai-review.md artifact

✅ No PR write permissions required

✅ No security risks

This turned out to be far more enterprise-compliant than auto-commenting.

🧪 Production Impact After enabling this system:

✅ Every PR is reviewed automatically

✅ Developers get feedback in minutes

✅ Review standards are enforced consistently

✅ Human reviewers focus only on business logic

✅ No failed pipelines due to auth issues

✅ No wasted build minutes on retries

✅ Zero per-developer licensing cost

🔑 Key Engineering Lessons
Avoid brittle platform APIs when Git can do the job

AI reviewers should assist, not block developers

PR comments are optional-reviews must be reliable

Pipelines + Git + LLM = extremely powerful combination

Groq is ideal for CI/CD AI workloads

Not every AI solution needs a $20/month/dev license

📌 What’s Next?
Planned upgrades:

✅ Auto-block merge when verdict = NEEDS WORK

✅ Language-specific reviewers (.NET, SQL)

✅ Security-only review mode

✅ Architectural drift detection

✅ Final Thoughts
If you're using Bitbucket and want reliable AI-powered PR reviews without paying enterprise per-seat pricing, my recommendation is:

💡 Use Git for diff extraction + Groq for AI analysis + Pipelines for automation. Avoid REST API auth wherever possible.

It’s simpler. It’s faster. It’s cheaper. And it actually works in production.