Linux / DevOps Engineer | HF Propagation Prediction | Sovereign AI | Amateur Radio (KI7MT, DN13)
IONIS-AI (Ionospheric Neural Inference System) — a self-hosted AI system that predicts HF radio propagation from one of the largest curated amateur radio datasets we are aware of. 14+ billion spots. Zero cloud dependencies.
The logs were speaking for decades, but nobody was listening. Now we're listening.
| Repo | Language | What It Does |
|---|---|---|
| ionis-apps | Go | High-performance data ingesters (22 Mrps) with watermark tracking |
| ionis-core | SQL/Shell | 36 ClickHouse DDL schemas, 15 population scripts |
| ionis-cuda | C++/CUDA | GPU-accelerated signature embedding engine |
| ionis-training | Python | PyTorch model training with physics-constrained sidecars |
| ionis-mcp | Python | MCP server for propagation analytics (11 tools, PyPI) |
| ionis-jupyter | Python | Jupyter notebooks for propagation research (10 notebooks, PyPI) |
| ionis-docs | MkDocs | Documentation site |
| Source | Volume | What It Tells Us |
|---|---|---|
| WSPR | 10.9B spots (18 yrs) | SNR floor — path exists at minimum power |
| Reverse Beacon Network | 2.26B spots (17 yrs) | Skilled operator layer — CW through the noise |
| CQ Contest Logs | 234M QSOs | Ceiling — voice-workable at contest power |
| PSK Reporter | ~26M spots/day (live) | Operational — can a real operator make a contact? |
| DSCOVR L1 | Solar wind (live) | Predictive — Bz gives 15-45 min lead over Kp |
IONIS V22-gamma (Production) — Pearson +0.49 vs VOACAP +0.02. 98% recall on independent live data. 207K params, physics-constrained solar and geomagnetic sidecars. Every watt of inference runs on local hardware.
Long-time Linux user (since the Slackware days) and active supporter of Open Source Software.
- Ubuntu user since 2005 (5.04), Ubuntu Member since ~2013
- RHEL/Fedora COPR packaging since ~2015
- Launchpad package maintainer since 2010
- Worked across Red Hat, Fedora, CentOS, Rocky, Arch, Gentoo, Alpine, and more
If it's Unix-y, I've probably broken it and fixed it.
Everything runs on local hardware. No cloud. No subscriptions. No vendor can revoke access.
| Host | Role | Specs |
|---|---|---|
| Threadripper 9975WX | Control node | 32C/64T Zen 5, 128 GB DDR5, RTX PRO 6000 96 GB VRAM, Rocky Linux 9.7 |
| Mac Studio M3 Ultra | Training node | 96 GB unified memory, MPS backend for PyTorch |
| EPYC 7302P | Build/replica node | 16C/32T, 128 GB DDR4 ECC, Proxmox hypervisor |
Networking: 10 Gbps DAC point-to-point (Thunderbolt 4 + SFP+ AOC), MTU 9000 Storage: ClickHouse on dedicated NVMe (3.7 TB), ZFS archive pool (7.1 TB mirrored) Data platform: ClickHouse (14B+ rows), PyTorch (MPS), Go ingesters (ch-go native protocol)
AI-agent micro-servers for Amateur Radio, all available on PyPI.
| Server | Tools | What It Does |
|---|---|---|
| ionis-mcp | 11 | Propagation analytics over 175M+ signatures, live NOAA conditions |
| adif-mcp | 7 | ADIF 3.1.6 validation, parsing, and spec search |
Open for discussions on Linux, DevOps, Amateur Radio, HF Propagation, or AI.
Reach me on GitHub Discussions or via ki7mt.io