← back to projects

AI-Augmented Development

A hybrid LLM workflow that plans on an expensive model and generates on a cheap one.

Results

What it did

70–85%cost cut by splitting reasoning from execution
2 modelsOpus plans, a cheaper model generates
ReviewedOpus checks the cheap model's output

The problem

Why it needed building

AI coding tools exist but most developers use them ad-hoc — a Copilot suggestion here, a ChatGPT query there. The real gains come from systematic integration: using the right AI model for each task, building custom workflows, and measuring actual productivity impact. The challenge was designing a repeatable system that delivers consistent 3-5x throughput across different project types.

The approach

How it works

Developed a hybrid architecture separating reasoning (Claude Opus for architecture and complex decisions) from execution (smaller models for code generation and tests). Custom workflows automate the full cycle: requirement analysis → architecture planning → code generation → test generation → review. Each phase uses the optimal model for cost-performance balance, achieving 70-85% cost savings versus using premium models for everything.

  1. Requirements
  2. Architecture
  3. Code Generation
  4. Test Generation
  5. AI Review
  1. 01
    RequirementsClaude Opus analyzes requirements and identifies edge cases
  2. 02
    ArchitectureClaude Opus plans structure, patterns, and interfaces
  3. 03
    Code GenerationGLM/Sonnet generates implementation from architecture
  4. 04
    Test GenerationAutomated test scaffolding with edge case coverage
  5. 05
    AI ReviewOpus reviews output for correctness, security, and style

Live demo

Try it yourself

Manual0 min
VS
AI-Assisted0 min

Tech stack

Built with

Claude CodeCursorCopilotCI/CDClaude OpusGLM-4GitHub Actions

Working on something like this?

Email me about itAll case studies