Hi everyone! I’m Gentian Elmazi, a software engineer with over ten years of experience and co-founder of Infinitcode.com and I am starting the first #buildinpublic series. At our company, we partner with businesses across industries—outsourcing the development of web applications powered by deep tech like AI, NLP, blockchain and so on.
Our Established Standards
From day one, we’ve enforced maintainable, readable, and scalable code. We adopt layered architectural patterns with dependency injection for modularity and testability. For source control, we follow GitHub Flow:
-
feature/*
branch - → Pull request
- → Peer review
- → Merge into
dev
- → Production release
The Bottleneck We Hit
As our client base and project scope grew, our senior engineers spent roughly 40% of their time on code reviews. Constantly context-switching across multiple repositories began to erode our strict standards—and that led to production issues we simply couldn’t ignore.
The Promise (and Pitfalls) of AI
When AI assistants promised 30–40% productivity boosts, we were eager to adopt them. Yet out-of-the-box AI snippets often:
- Violated our company rules
- Introduced inconsistent patterns
- Missed critical edge cases
Piloting Infinitcode.ai’s Code Reviewer
We decided to take a leap of faith and built an internal AI code-reviewer—now in public alpha as Infinitcode.ai—to help streamline our workflow. Within days, we saw:
35% Productivity Gain
Automated PR summaries and inline suggestions let senior engineers focus on high-value reviews, shrinking review cycles from days to hours.30% Performance Improvement
The AI flagged inefficient loops and unnecessary computations that we’d normally catch only after profiling in production.15+ Security Bugs Caught
Critical vulnerabilities surfaced early in pull requests, not weeks after deployment.120+ Typos & Style Violations Fixed
Consistent code formatting and improved documentation boosted overall readability.
Example Security Fix
⚠️ Risk: Non-cryptographic
UUIDv4
generation in bulk operations risked collisions under high-concurrency loads.
🔧 Fix: Implementedcrypto.randomUUID()
with batch-safe collision checks, ensuring unique identifiers even when processing thousands of records per second.
Even with a decade of hands-on coding, I likely would’ve missed this subtle bug until it hit production—so having AI catch it early was truly a game-changer.
The charts below illustrate the 35% productivity and 30% performance improvements we achieved—plus the dozens of security bugs and 120+ typos our AI reviewer caught—demonstrating its measurable impact on our workflow
Key Features That Helped Us
Multi-Model Integration
DeepSeek’s backbone matched our use cases, letting us switch models as needed.Custom Rulesets
Uploading our linting and security policies ensured AI suggestions aligned perfectly with our coding standards.Rapid Support & Iteration
Working directly with the team allowed us to fine-tune the tool within hours, not weeks.
Lessons Learned & Next Steps
AI Empowers, Doesn’t Replace
Developers remain central—AI handles routine checks so humans tackle design and architecture.Iterate on Your Rules
Continuously refine your AI’s rulesets and model settings as your codebase evolves.Measure & Visualize
Track metrics (productivity gains, bug catch rate, review times) to demonstrate ROI and keep stakeholders aligned.
I’d love to hear how AI is reshaping your code reviews—share your experiences or questions in the comments below!
As well feel free to try Infinitcode.ai in alpha stage. Your feedback is greatly appreciated
That's interesting. I'd like to talk more about the processes that your AI is doing. How can you ensure that AI can produce code that is really "necessary"? I've used different AIs and the code they were producing had added complexity and I think that's a problem that you and your company will soon face. A lot of code that has added complexity. And it just creates additional overhead.
For example, if you want to generate a simple crud project with clude, he will probably write some extra code, for example, he may do extra work for a simple sum, while we only want to have a sum function.
Anyway, I don't think AI should be introduced into the code anytime soon. He should be responsible for correcting "spelling errors".