AI Code Generation System: 6x Faster Development

We often encounter a situation where a team of 4–5 developers spends 4–6 hours creating a single standard CRUD endpoint with tests. Meanwhile, the business demands 3–5 new endpoints per week — routine takes 70% of the time, leaving little for business logic and architecture. This is the problem solved by an AI code generation system: it takes over the boilerplate, leaving the engineer with creative tasks.

How AI Code Generation Accelerates Development

A custom AI agent doesn't just insert code from a template — it understands the context of your codebase: DB schemas, existing classes, API contracts, code style. Based on that, it generates production-quality code, checks syntax, runs tests, and iteratively fixes errors. Architecturally, such a system includes:

• Context Manager — collects relevant context: DB schema, API interfaces, existing models, code style guide.
• Generation Engine — LLM agent with tools for reading files, running tests, searching the codebase.
• Verification Layer — syntax checking, test execution, linter.
• Feedback Loop — iterations based on test errors.

from langgraph.graph import StateGraph, END
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from typing import TypedDict, Annotated, Optional
import subprocess
import ast
import operator

llm = ChatOpenAI(model="claude-opus-4-5", temperature=0.1)

class CodeGenState(TypedDict):
    task_description: str
    existing_code_context: str
    generated_code: Optional[str]
    test_results: Annotated[list, operator.add]
    iteration: int
    max_iterations: int
    errors: Annotated[list, operator.add]
    final_code: Optional[str]

@tool
def read_file(file_path: str) -> str:
    """Read a file from the codebase to get context."""
    try:
        with open(file_path) as f:
            return f.read()
    except FileNotFoundError:
        return f"File {file_path} not found"

@tool
def search_codebase(query: str, directory: str = "./src") -> str:
    """Search the codebase with grep to find similar code."""
    result = subprocess.run(
        ["grep", "-r", "--include=*.py", "-n", query, directory],
        capture_output=True, text=True
    )
    return result.stdout[:3000] if result.stdout else "Nothing found"

@tool
def run_python_syntax_check(code: str) -> str:
    """Check Python code syntax."""
    try:
        ast.parse(code)
        return "Syntax is correct"
    except SyntaxError as e:
        return f"Syntax error: {e}"

@tool
def run_tests(test_file_path: str) -> str:
    """Run pytest and return results."""
    result = subprocess.run(
        ["python", "-m", "pytest", test_file_path, "-v", "--tb=short"],
        capture_output=True, text=True, timeout=60
    )
    output = result.stdout + result.stderr
    return output[-3000:]  # Last 3000 characters

@tool
def write_file(file_path: str, content: str) -> str:
    """Write code to a file."""
    with open(file_path, "w", encoding="utf-8") as f:
        f.write(content)
    return f"File {file_path} written ({len(content)} characters)"

CODE_GEN_SYSTEM = """You are a Senior Software Engineer. Generate production-quality code.

Principles:
- Follow existing codebase patterns
- Write typed code (type hints)
- Each function has one level of abstraction
- Handle errors explicitly
- Minimize dependencies on external libraries if standard alternatives exist

Process:
1. Read existing code for context
2. Generate code in the same style
3. Check syntax
4. Run tests
5. Fix errors iteratively"""

from langgraph.prebuilt import create_react_agent

code_gen_agent = create_react_agent(
    llm.bind_tools([read_file, search_codebase, run_python_syntax_check, run_tests, write_file]),
    tools=[read_file, search_codebase, run_python_syntax_check, run_tests, write_file],
    state_modifier=CODE_GEN_SYSTEM,
)

Why Codebase Context Matters

Without context, LLMs generate code that doesn't fit into the existing architecture — different style, wrong names, incompatible imports. Our Context Aware Code Generator automatically collects relevant files: data models, base classes, code style guide. This is critical for projects using FastAPI + SQLAlchemy with custom patterns. Example implementation:

class ContextAwareCodeGenerator:

    def __init__(self, project_root: str):
        self.project_root = project_root
        self.context_cache = {}

    async def gather_context(self, task: str) -> str:
        """Gather relevant context for the task"""

        # Find similar files via LLM
        relevant_files = await self.identify_relevant_files(task)

        context_parts = []

        # Read DB schema
        if await self.file_exists("models.py"):
            models = await read_file_async(f"{self.project_root}/models.py")
            context_parts.append(f"## Data Models\n{models[:2000]}")

        # Read base classes and interfaces
        for file_path in relevant_files[:3]:
            content = await read_file_async(file_path)
            context_parts.append(f"## {file_path}\n{content[:1500]}")

        # Add code style guide
        if await self.file_exists(".codestyle.md"):
            style = await read_file_async(f"{self.project_root}/.codestyle.md")
            context_parts.append(f"## Code Style\n{style[:1000]}")

        return "\n\n".join(context_parts)

    async def generate(self, task: str, output_file: str) -> dict:
        context = await self.gather_context(task)

        result = await code_gen_agent.ainvoke({
            "messages": [{
                "role": "user",
                "content": f"""Task: {task}

Codebase context:
{context}

Output file: {output_file}

Generate the code, check it, and write to file."""
            }]
        })

        return {
            "task": task,
            "output_file": output_file,
            "iterations": result.get("iteration", 1),
            "tests_passed": self.extract_test_status(result),
        }

Template-based Generation with LLM Filling

For typical tasks (CRUD, migrations, tests), a hybrid approach is effective: a template with placeholders that the LLM expands. This provides predictable structure and control over critical parts.

class CRUDGenerator:
    """Generates CRUD modules from entity schema"""

    CRUD_TEMPLATE = """
# Module for entity {entity_name}
from sqlalchemy import Column, Integer, String, DateTime, func
from sqlalchemy.orm import Session
from pydantic import BaseModel
from typing import Optional, List
from datetime import datetime

# PLACEHOLDERS FOR LLM REPLACEMENT:
# COLUMNS - list of SQLAlchemy columns
# PYDANTIC_FIELDS - Pydantic schema fields
# BUSINESS_LOGIC - specific business logic
"""

    async def generate_crud_module(self, entity_spec: dict) -> str:
        """entity_spec: {name, fields, business_rules, relationships}"""

        # LLM fills specific parts
        columns = await self.generate_sqlalchemy_columns(entity_spec["fields"])
        schemas = await self.generate_pydantic_schemas(entity_spec["fields"])
        business_logic = await self.generate_business_logic(entity_spec.get("business_rules", []))

        # Assemble final module
        result = await llm.ainvoke(f"""Create a full CRUD module for entity {entity_spec['name']}.

Specification: {json.dumps(entity_spec, ensure_ascii=False)}

Stack: FastAPI + SQLAlchemy 2.0 + Pydantic v2
Include: model, pydantic schemas, CRUD functions, FastAPI router with dependency injection
Code standards: async/await, type hints, docstrings""")

        return result.content

Case Study: Fintech Startup

Our client — a fintech company with 4 developers — spent 4–6 hours on a standard CRUD endpoint with tests. After implementing AI generation, the time dropped to 50 minutes (15 min generation + 35 min review). Metrics before and after:

The situation allowed the team to save significant costs on routine tasks, and the cost per PR decreased several times. The system generated CRUD modules from OpenAPI specifications, automatically created pytest tests and Alembic migrations. An AI Code Review agent provided suggestions at the review stage. The only challenge was business logic in 14% of cases requiring substantial rework, which is addressed by adding rules to the context.

Model Comparison for Code Generation

Model selection depends on budget and quality requirements. For production tasks, we recommend Claude 3.5 Sonnet or GPT-4o with large context window.

What's Included

When ordering our service, you receive:

• Codebase and architecture audit — assessment of CI/CD maturity, code style, test coverage.
• Design and implementation of an AI agent — tailored to your stack and requirements.
• Integration with IDE (VS Code, JetBrains) and CI/CD (GitHub Actions, GitLab CI, Jenkins).
• Team training — workshops on using the system, best practices for prompting.
• Documentation and 1 month warranty support.

Order AI generation implementation and accelerate your development.

Estimated Timelines

• Basic generator with context: 2–3 weeks
• Agentic loop with tests and iterations: 2–3 weeks
• Integration into CI/CD and IDE: 2–3 weeks
• Total: 6–9 weeks

Pricing is calculated individually, based on scope and complexity. Contact us for a project assessment — we'll propose the optimal solution. Our LangGraph agent is already running in several production systems, we guarantee stability and support.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.