Athena — AI command centerLive
A private, always-on console that triages email across inboxes, monitors AI and breach news, tracks weather, and runs the company back office — with four interchangeable LLM providers behind one router.
Athena is the private "glass cockpit" for Lockman Cyber. It runs on a Mac mini at the edge of the lab and is only reachable over a private mesh network. It is also the clearest example of how the company uses language models: as replaceable components behind a stable interface, with local-first defaults and cloud fallbacks.
What it does
- Email triage across inboxes. Every connected mailbox is scanned on a schedule. A model reads the new mail and surfaces only what needs a reply or an action, then the important FYIs. Promotions, receipts, and social noise never make the board. A "Done" button acknowledges an item so it never returns, and a "Not important" button teaches the prompt what to filter next time.
- News and breach monitor. RSS from technical and general AI sources plus security outlets is bucketed by a model into three boards. The breach board is a rolling 30-day view of real, named incidents, with a 24-hour alert bar when something new lands.
- Weather and tropical activity. Open-Meteo forecasts and NHC storm data, with acknowledgeable storm alerts that re-fire if a system strengthens.
- Back office. Invoicing, expense tracking, mileage, home-office proration, and a record book for the LLC, all with server-side persistence, wipe guards, and one-deep backups.
- Status board. Health of the internet link, the mini, the Jetson, the hypervisor, the trading bot, and the upstream model providers.
- Kiosk mode. A touch layout for a wall-mounted panel, with a live radar tile and big tap targets. See the touchscreen status panel.
The interesting part: the brain router
Every feature calls one function, callBrain(prompt, feature), and a settings screen decides which provider answers for each feature. Four brains are wired in for real:
| Brain | Where it runs | Used for |
|---|---|---|
| Local Ollama | NVIDIA Jetson Orin Nano on the LAN | Free, private default and fallback |
| Ollama Cloud | gpt-oss, Kimi, DeepSeek, Qwen | Fast general judgment |
| OpenAI | GPT-5 family | JSON-mode classification |
| Anthropic | Claude Opus, Sonnet, Haiku | Sharp triage where judgment matters |
Switching is on the fly. Prompts force JSON and a tolerant parser cleans up fenced or chatty responses. Small local models get tuned context windows and smaller batches; the notes on which cloud models were fast, accurate, or retired were all learned empirically and kept in the repo.
Lessons that shaped later client work
- Sticky merges beat re-judging. Models re-judge the same inbox slightly differently each pass. Surfaced items now persist until a human acknowledges them or the source confirms they are gone.
- Coverage bugs hide as "the AI missed it." An important email was never fetched because the collector capped at the newest N messages. The fix was incremental scanning plus a visible "triaged N of M" counter so a silent cap can never hide again.
- Deploy atomically. The wall panel reloads the moment the page changes, so a half-written deploy once killed it. Everything ships via write-then-rename now.
Athena itself is never exposed publicly. Only the architecture is.