How a code graph saves 80% of your AI tokens

Your AI coding assistant is spending most of its tokens just searching through your files. What if there was a smarter way?

In this episode, Romain sits down with Davide de Sio, Head of Software Engineering at Eleva and AWS Community Builder, to discuss KiroGraph — a 100% local semantic code knowledge graph that turns your codebase into a queryable graph structure, dramatically reducing AI tool calls and token usage.

🔑 Key topics covered:
- Why AI agents waste tokens on file search (and how a code graph fixes it)
- Tree-sitter: the AST technology powering KiroGraph's code understanding
- The Architecture module: preventing AI-induced architecture drift
- The Security module: detecting exposed API keys and vulnerabilities in the graph
- Watchman: auto-generating Kiro skills from repetitive developer patterns
- Running everything locally with Nomic embeddings and Gemma 3 (no cloud dependency)
- Portability: carrying your code graph across machines via Git
- Future: CI/CD integration, validation loops with hooks, and container deployment

⏱️ Timestamps:
00:00 - Introduction & Davide's background
02:50 - What is KiroGraph and why he built it
06:24 - How token/credit consumption works in Kiro
08:25 - Code graphs vs. file search: a smarter approach
09:41 - Tree-sitter and AST-based graph generation
11:23 - "Building for Kiro with Kiro" — the inception story
12:12 - Architecture module deep-dive
18:28 - How KiroGraph wires into Kiro (hooks, steerings, skills)
22:47 - Validation loops with hooks (spoiler: new module coming)
25:12 - Persistent memory & the Watchman module
27:33 - Why everything runs 100% locally
30:00 - Cross-platform support and community contributions
32:00 - Local models: Nomic for embeddings, Gemma for summarization
35:40 - The hybrid future: local + cloud AI
37:21 - Security module & Caveman integration
43:52 - Portability between machines
47:52 - What's next: community-driven roadmap
54:51 - Token savings data: up to 80% credit reduction
59:12 - Closing: community-driven development & AWS Summit story

🔗 Resources mentioned:
- KiroGraph (GitHub): https://github.com/davide-desio-eleva/kirograph
- Kiro IDE: https://kiro.dev
- Tree-sitter: https://github.com/tree-sitter/tree-sitter
- Nomic Embed (Hugging Face): https://huggingface.co/nomic-ai/nomic-embed-text-v1
- AWS Community Builders: https://aws.amazon.com/developer/community/community-builders/
- Davide's blog post: https://devs.30tools.com/aws-builders/building-kirograph-a-100-local-semantic-code-knowledge-graph-for-kiro-2ja4

👤 About the guest:
Davide de Sio is Head of Software Engineering at Eleva, based near Milan, Italy. He's an AWS Community Builder specializing in serverless architectures and the creator of KiroGraph.
LinkedIn: https://www.linkedin.com/in/desiodavide/

🎙️ About the AWS Developers Podcast:
Weekly conversations about Agentic AI and the future of software development with customers, AWS community members, and industry luminaries. Hosted by Romain Jourdan.

📩 Subscribe & Follow:

Видео How a code graph saves 80% of your AI tokens канала AWS Developers Podcast