My Friend Was Tired of AI Answers That Sounded Right. So I Built One That Checks Itself.

Pooja — Sun, 04 Oct 2026 21:45:31 +0000

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

I built ARIA (Adaptive Retrieval Intelligence Architecture) for a friend who is studying cybersecurity.The problem was pretty simple: when he was studying topics like SQL injection, OWASP, or MITRE ATT&CK, he often had to go through a lot of different documents to find the exact information he needed. Even when an AI gave him a correct answer, it didn't always match the material he was actually studying.

So I built ARIA to make that process easier.

ARIA lets him add his own cybersecurity resources and ask questions based on those materials. Instead of relying only on what the language model already knows, it retrieves the most relevant information from his knowledge base and uses that context to answer the question.

I also wanted it to handle situations where the first retrieval isn't good enough. ARIA checks the retrieved context and can try to improve the retrieval instead of simply giving an answer based on weak or unrelated information.

The goal wasn not to build another general-purpose chatbot. I wanted to build something my friend could actually use while studying — something that understands the resources he is working with and answers from them.

Demo

Here is the Deployment link :
https://huggingface.co/spaces/miniorange/aria-cybersecurity

Code

Dear frnd , check my github : ARIA FOR YOU

How I Built It

ARIA is built around open-source and locally runnable AI. I use Llama 3.1 8B through Ollama for generation, with ChromaDB and BM25 for hybrid retrieval. The system retrieves relevant cybersecurity knowledge, then uses a Judge LLM to evaluate the retrieved context and the generated answer. If the retrieval or answer isn't good enough, ARIA's self-healing engine can refine the retrieval and try again.

The whole idea was to keep the pipeline transparent and modular, so the models and retrieval components can be inspected, evaluated, and replaced without rebuilding the entire system.

Why Does Open Innovation Matter?

Open innovation made ARIA possible because I could build the entire retrieval and evaluation pipeline around models and tools that I can actually inspect and control.

Instead of sending my friend's cybersecurity study material to a closed API and treating the model's response as a black box, ARIA can keep the knowledge base and inference workflow under the user's control. I can also experiment with different models, retrieval methods, evaluation techniques, and self-healing strategies without being locked into one provider.

For a project focused on grounded and trustworthy answers, that control really matters.

Prize Categories

GitHub Copilot: Used GitHub Copilot throughout the development of ARIA to help with implementation, debugging, and iterating on the retrieval and self-healing pipeline.