lablab-ai-amd-developer-hackathon/helios

HELIOS — Real-Time Space Weather Intelligence

First multi-agent AI system for real-time solar storm detection and infrastructure impact forecasting. Running NASA/IBM Surya-1.0 on AMD Instinct MI300X via ROCm.

AMD Developer Hackathon · May 4–10, 2026 · lablab.ai

Demo

🎥 https://www.loom.com/share/a15bf3391c0945e7950ff213460d3ced  Live dashboard: http://134.199.197.132 (active during hackathon)

What It Does

HELIOS watches the Sun 24/7, detects solar storms as they form, models how they travel through space, and delivers plain-language impact forecasts to operators — telling them exactly which satellites, power grids, GPS systems, and aviation routes face risk, and when.

The problem: When a CME arrives at Earth, the DSCOVR sensor at the L1 Lagrange point gives operators approximately 30 minutes of real-time warning — verified at 31 minutes for the May 2024 Gannon G5 storm. That is not enough time to complete protective actions for critical infrastructure.

HELIOS detects flares at the solar source using GOES X-ray data and NASA's Surya foundation model, issuing automated WARNING alerts days before Earth impact — rather than waiting for the storm to arrive at DSCOVR's L1 position. For the Gannon storm, our pipeline issued a WARNING 36 hours before impact, validated against real NASA DONKI and GFZ Potsdam archives.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    LIVE DATA SOURCES                    │
│  NASA SDO (images)  │  NOAA DSCOVR  │  GOES X-ray      │
└──────────┬──────────┴───────┬────────┴──────┬───────────┘
           │                  │               │
           ▼                  ▼               │
  ┌─────────────────┐  ┌─────────────────┐   │
  │   AGENT 01      │  │   AGENT 02      │◄──┘
  │  Solar Vision   │  │  CME Physics    │
  │  Surya-1.0      │  │  DSCOVR L1      │
  │  GOES X-ray     │  │  Burton / DBM   │
  └────────┬────────┘  └────────┬────────┘
           │    (parallel)      │
           └────────┬───────────┘
                    ▼
          ┌─────────────────┐
          │   AGENT 03      │
          │ Impact Mapper   │
          │  Kp → Infra     │
          │  Folium map     │
          └────────┬────────┘
                   ▼
          ┌─────────────────┐
          │   AGENT 04      │
          │  Command LLM    │
          │  Llama 3.1 8B   │
          │  Alert bulletin │
          └────────┬────────┘
                   ▼
          ┌─────────────────┐
          │ OPERATOR DASH   │
          │  Streamlit UI   │
          │  Live + Replay  │
          └─────────────────┘

Agents 01 and 02 run in true parallel via ThreadPoolExecutor — GOES/SDO fetch and DSCOVR fetch are independent. They sync at Agent 03.

Benchmark — Surya-1.0 on AMD MI300X

Why MI300X: Llama 3.1 8B (~16 GB fp16) + Surya (~1.82 GB) + vLLM overhead — all loaded simultaneously in one VRAM pool. No model swapping. The MI300X's 5.2 TB/s memory bandwidth enables real-time inference within SDO's 12-second image cadence.

Tech Stack

Models used in this submission: Surya-1.0 (366M) + Llama 3.1 8B Instruct via vLLM on AMD ROCm 7.2.

Data Sources

All data is free, public, and streaming in real time:

Agent Specifications

Quickstart

Fresh AMD Developer Cloud instance — complete setup in one command per step:

# On the AMD host (run once):
bash <(curl -s https://raw.githubusercontent.com/hadsaw-parallel/helios/main/setup_host.sh)

# Inside the Docker container:
docker exec -it rocm /bin/bash
cd /app && git clone https://github.com/hadsaw-parallel/helios.git && cd helios && bash setup.sh

setup.sh automatically:

1. 1.Clones the repo and installs dependencies
2. 2.Clones NASA-IMPACT/Surya and installs it
3. 3.Downloads Surya-1.0 weights from HuggingFace
4. 4.Starts vLLM serving Llama 3.1 8B
5. 5.Starts Streamlit dashboard on port 30000
6. 6.Proxied to port 80 via Caddy

Dashboard opens at `http://<YOUR_IP>` (~15 min from zero on a fresh instance).

Running Tests

python3 -m pytest tests/ -v
# 9 passed, 1 skipped (storm replay requires SuryaBench data)

What Makes HELIOS Unique

1. 1.First demonstration of NASA/IBM Surya-1.0 on AMD ROCm hardware. Surya's GitHub targets CUDA only. HELIOS ports and runs it on MI300X — a direct AMD ecosystem contribution.

1. 1.First multi-agent agentic pipeline for operational space weather forecasting. Existing tools are siloed: separate apps for solar imaging, solar wind data, and impact assessment. HELIOS chains them into one autonomous pipeline.

1. 1.Scientifically grounded physics. Agent 02 uses real DSCOVR measurements (not synthetic data) and the Burton (1975) empirical formula for Kp estimation. Agent 03's latitude bands match NOAA's published G-scale.

1. 1.Validated against real storms. The counterfactual replay fetches live data from NASA DONKI, GFZ Potsdam Kp API, and NASA OMNIWeb for any historical timestamp — nothing hardcoded. Validated against the May 2024 Gannon G5 storm: HELIOS issued WARNING at T-36h using real archived flare data. The live pipeline shows real conditions (ALL_CLEAR when the Sun is quiet).

The Warning Window

HELIOS detects X-class flares at the solar source using GOES X-ray — the same data NOAA analysts use, but in an automated pipeline that also maps infrastructure impact and generates operator bulletins in under 3 seconds. DSCOVR provides the final real-time confirmation as the storm arrives; HELIOS provides the early automated alert before it does.

Validated Claim

"The May 2024 Gannon G5 storm — strongest in 21 years. DSCOVR gave operators 31 minutes of real-time warning when the CME was already arriving. HELIOS, running the same public GOES data through an automated pipeline, issued a WARNING 36 hours before impact — validated live against NASA DONKI and GFZ Potsdam archives."

Built with NASA/IBM Surya-1.0 · AMD Instinct MI300X · ROCm 7.2 · LangGraph · Llama 3.1 AMD Developer Hackathon · May 4–10, 2026