Week #1 #

Project description #

Project name: Scaffold #

Code repository:

• Course Repository
• Our Repository

Scaffold is a temporary structure used to support a work crew and materials to aid in the construction, maintenance and repair of buildings

Our Scaffold is a code management system designed to translate and maintain source code as a graph in a graph database, enabling seamless context injection for large language models (LLMs). And helps AI agents in construction, maintenance and repair of your project.

Team Members #

Brainstorming #

Ideas during brainstorming #

1 Graph-based code context platform for LLMs Translate source code into a graph database (AST, function/class relations) to serve as rich structured context for AI agents. Enables scalable, accurate retrieval of relevant information for code generation and QA.

2 AI codebase companion (Scaffold CLI) CLI tool integrated into developer workflows that allows querying, summarizing, or modifying the codebase using LLMs with graph-backed context.

3 LLM-aware refactoring assistant Leverages code graphs and embeddings to propose or automate safe refactoring operations (rename symbols, split/merge functions, remove dead code).

Brief market research / problem validation #

Idea 1: Graph-based code context platform for LLMs Problem: Modern LLMs operate on tokenized text and lack awareness of the structural and semantic organization of real-world codebases. Existing solutions (e.g., embedding chunks into a vector DB) do not capture hierarchical or reference-based relationships well.

Existing solutions: Tools like Sourcegraph Cody, Codeium, and GitHub Copilot use text embeddings but struggle with large-scale project structure and maintaining long-term context.

Validation: Research from OpenAI, Meta, and others highlights the importance of hierarchical and symbolic context in improving AI performance on large-scale code reasoning tasks. Graph-based representations are also used in tools like CodeQL for similar reasons.

Idea 2: LLM-aware refactoring assistant Problem: Refactoring at scale (e.g., renaming a core service method used in hundreds of files) is high-risk and hard to reason about, especially across language boundaries.

Existing solutions: IDEs like IntelliJ or VSCode offer local static analysis refactors, but not AI-assisted reasoning or graph-level semantic refactoring.

Validation: Enterprise engineering teams report significant friction in large-scale refactoring, especially when team members are unfamiliar with legacy code or there’s poor documentation. GitHub Copilot lacks this structured reasoning.

Basic requirements #

Parse code into AST and build code graphs

Store in a graph DB (e.g., Neo4j) and vector DB (e.g., Qdrant)

Extract structural/code entity relationships (calls, imports, etc.)

Provide API/CLI for context queries

Support incremental updates (e.g., Git hooks or file watchers)

Enable context injection into LLMs (RAG)

Basic testing and validation tools

Target users and their primary needs #

Developers Understand and refactor code faster using AI and graph context AI Engineers Provide structured context to LLMs for better accuracy Tech Writers Auto-generate or update documentation from code structure QA Engineers Understand dependencies and test impact of code changes

User stories #

As a developer, I want to find all references to a function to safely rename it.

As an AI engineer, I want structured code context to improve RAG results.

As a tech writer, I want to auto-generate docs from code relationships.

As a QA engineer, I want to trace service dependencies for better test coverage.

Initial scope #

Python code parser → graph + vector DB

Neo4j + Qdrant integration

Basic API/CLI for context lookup

LLM context injection (early RAG prototype)

CLI tool for developers

Basic graph update system (e.g., file watcher)

Tech-stack #

Python – Widely used in AI and tooling; ideal for building parsers, integrating LLMs, and rapid prototyping.

Neo4j – Purpose-built graph database optimized for modeling and querying complex code relationships.

VectorDB (e.g., Qdrant) – Enables high-performance semantic search over embedded code/document chunks.

Docker – Provides consistent, containerized environments for development, testing, and deployment.

LLM Chain (e.g., LangChain) – Modular framework for orchestrating Retrieval-Augmented Generation pipelines.

Weekly commitments #

Individual contribution of each participant #

Melnikov Sergei - brainstorming, repository, informtaion research Razmakhov Serhei- brainstorming, repository Prosvirkin Dmitry - brainstorming, writing report Mashenkov Timofei - brainstorming, informtaion research Glazov Sergei - brainstorming, writing report

Confirmation of the code’s operability #

We confirm that the code in the main branch:

• In working condition.
• Run via docker-compose (or another alternative described in the README.md).