explaingit

srkrz23/atlas

0PythonAudience · developerComplexity · 4/5ActiveLicenseSetup · moderate

TLDR

Hackathon multi-agent system that takes voice input, screens it for jailbreaks, routes subtasks to open source models, and writes a signed audit chain of every step.

Mindmap

mindmap
  root((ATLAS))
    Inputs
      Voice
      Speechmatics transcript
    Outputs
      Final answer
      Signed audit log
    Use Cases
      Compliant AI assistant
      Prompt injection screening
      Multi-model routing
    Tech Stack
      Python
      Streamlit
      Gemini
      Featherless
      Ed25519

Things people build with this

USE CASE 1

Demo a compliance-friendly multi-agent stack with a signed audit trail

USE CASE 2

Route different task types to MiniMax, DeepSeek, Kimi, or Llama 3.3 70B automatically

USE CASE 3

Reject prompt injection and jailbreak attempts before they reach the planner

USE CASE 4

Pull live Kraken or Vultr API data through agent tool calls

Tech stack

PythonStreamlitGeminiFeatherlessSpeechmatics

Getting it running

Difficulty · moderate Time to first run · 30min

Runs in mock mode without keys, but full features need Speechmatics, Gemini, Featherless, Kraken, and Vultr credentials.

MIT permits commercial use, modification, and distribution with attribution.

In plain English

ATLAS is a hackathon project submitted to the Milan AI Week Hackathon 2026. It is a multi-agent system, meaning it strings several AI components together into one pipeline, and its main selling point is that every step in that pipeline is checked, logged, and cryptographically signed so that an enterprise could later inspect what the system actually did. The flow starts with voice input that is transcribed by Speechmatics. The text then passes through a layer called SOUF AI DPI, which inspects the prompt for things like jailbreak attempts, hidden instructions, base64 tricks, or unusual unicode characters. If the prompt looks adversarial, it is rejected with a 403 response and a regulation citation. Otherwise it moves on to a Gemini 2.0 Flash model that acts as the planner and breaks the request into smaller subtasks. Each subtask is then handed to a router that picks an open source model from Featherless based on what kind of work is needed. Agent style tool use goes to MiniMax, code work goes to DeepSeek, long context and multilingual work goes to Kimi, and general questions go to Llama 3.3 70B. The chosen model can call tools, which in this project include DuckDuckGo search, a read only sandbox database, the Kraken API for live crypto market data, and the Vultr API for infrastructure queries. Every action the system takes, from the initial prompt check through tool calls and the final answer, is recorded into an audit log. Each entry is signed with an Ed25519 key and chained to the previous entry with a SHA-256 hash, so changing any earlier record breaks the chain. Gemini also produces a final synthesized answer at the end of the pipeline. The README explains how to run it. A pip install plus a streamlit command launches the user interface, and the system works in mock mode without any API keys, which is useful for demos. There is a test suite the author reports as 29 out of 29 passing, a short video walkthrough, a pitch deck, and a project structure listing the modules for the gateway, orchestrator, router, tools, audit chain, and frontend. The license is MIT.

Copy-paste prompts

Prompt 1
Install ATLAS with pip, launch Streamlit in mock mode, and run a voice prompt through the pipeline
Prompt 2
Add a new tool that queries the GitHub REST API and register it with the Featherless router
Prompt 3
Verify the Ed25519 signed audit chain by tampering with one log entry and confirming the next break
Prompt 4
Swap Gemini 2.0 Flash for Claude Sonnet as the planner and re-run the test suite
Prompt 5
Add a new SOUF AI DPI rule that blocks prompts containing more than three base64 blobs
Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.