Skip to content
@Yggdrasil-AI-labs

Yggdrasil AI Labs

Applied AI engineering, self-hosted and local-first. RF-to-data tooling plus the agent fleet that runs the lab. Built to find the limits.

Yggdrasil AI Labs

Self-hosted AI operations & signals (SIGINT) tooling: open-source tools that turn real-world RF into clean, structured data, plus the disclosure tooling that keeps publishing safe. All on gated CI/CD.

Open-source tools

Built and run in the lab: human-directed, AI-assisted development, gated CI/CD, and the agent fleet below that operates the infrastructure. Three families.

Signal pipeline: capture, normalize, transport

Repo What it does
adsb-to-wdgwars Normalizes 13 ADS-B receiver dialects into one JSON schema · CLI + in-browser (Pyodide/WASM)
meshcore-to-wdgwars LoRa / MeshCore mesh telemetry → normalized records
wigle-to-wdgwars WiGLE Wi-Fi / BLE wardrive CSVs → structured records
gungnir Shared HMAC-signed transport client: integrity, retry, cooldown, silent-drop detection
wdgwars-api-tester Contract-testing harness for an undocumented HTTP API: probe quorum + 404 fingerprinting

Readout: consume, display

Repo What it does
wdgwars-discord-stats Read side of the same API the feeders push to: live WDGoWars stats as a Discord voice-channel dashboard or webhook post · stdlib-only, read-only

Disclosure tooling

Repo What it does
predisclose Pre-publish disclosure scanner: secrets, PII, and internal identifiers. Stdlib-only core; generic patterns public, your inventory stays local

How it's built

  • One transport, many tools: signing, retry, and back-pressure live once in gungnir; the tools stay small and focused.
  • Local-first: browser builds process data in-page (Pyodide/WASM); no server sees raw input.
  • Gated CI/CD on every repo: pytest + coverage → SonarCloud quality gate (SAST) → Snyk SCA → release artifact, enforced by branch protection on main.
  • MIT-licensed, no telemetry.

AI Ops

Yggdrasil AI Labs is a self-hosted lab for applied AI engineering: building local-first systems end to end, then running them hard enough to find the limits. The signal-pipeline tools above turn real-world RF into structured data. The lab itself is operated by a fleet of around twenty small, single-purpose agents that handle the work a person would otherwise do by hand, so the engineering effort goes into building and stress-testing, not babysitting.

The design is local-first. Agents run inference on local GPUs and only fall back to a hosted model when a task needs it, which keeps recurring work cheap and on-prem. Each agent is a small, dependency-light service that writes its own state and journal, fails safe, and reports to a shared dashboard, so the fleet stays observable and any one piece can be reasoned about in isolation.

Do the work: Seer (image understanding + OCR), Scribe (speech-to-text), Janitor (knowledge-base hygiene), Librarian (canonical writing + cross-linking), Secretary (documentation-coverage audit), Recruiter (proposes a new agent when recurring work justifies one), Triage (classifies + routes incoming notes), Researcher (drains a research queue into synthesis notes).

Run the lab: Heimdall (job orchestrator with concurrency limits + priorities), Doctor (scheduled health checks with safe auto-fixes), Forseti (the local model that reasons over Doctor's findings), Probe (integration smoke tests across every component), Vidar (remote-host + mesh-node reachability with an expected-offline allowlist).

Watch the lab: Bursar (per-agent cost + local-vs-cloud routing), Quartermaster (local model + GPU-fit management), Evaluator (a fixed eval set that gates local-model quality over time), Sentinel (defensive posture checks on the lab's own surfaces), Norn (configuration-drift detection against a declared baseline), Herald (a daily rollup brief), Cartographer (regenerates the live architecture map from real state).

Together they form a loop: workers do tasks on the local-first tier, everything writes to a shared state layer, and the watchers read it back to keep the fleet cheap, correct, and current. Same discipline as the feeders: small single-responsibility services, local-first compute, gated changes, no telemetry.

Popular repositories Loading

  1. adsb-to-wdgwars adsb-to-wdgwars Public

    Muninn: ADS-B feeder for WDGoWars. Handles AVR, SBS-1, dump1090/readsb, tar1090, VRS, Stratux, Mode-S Beast, NDJSON, Mayhem, GDL-90, and CSV. Built on gungnir (shared transport). Ships as both CLI …

    Python 16

  2. wigle-to-wdgwars wigle-to-wdgwars Public

    Push WiGLE-format Wi-Fi/BLE wardrive CSVs to the WDGoWars leaderboard. Built on gungnir (shared transport). Companion to Muninn (adsb-to-wdgwars) and Heimdall (meshcore-to-wdgwars).

    Python 8

  3. meshcore-to-wdgwars meshcore-to-wdgwars Public

    Heimdall: MeshMapper Meshcore LoRa feeder for WDGoWars, targeting the meshcore_nodes payload slot. Ships as both CLI (Python) and browser (Pyodide) with the same parser core.

    Python 5

  4. gungnir gungnir Public

    Shared transport client for the WDGoWars (wdgwars.pl) feeders: HMAC envelope, retry, cooldown, silent-drop detection. Used by Muninn, Heimdall, wigle-to-wdgwars.

    Python 2

  5. wdgwars-discord-stats wdgwars-discord-stats Public

    Build your own WDGoWars stats display in Discord: live voice-channel dashboard, webhook poster, and a consolidated API reference.

    Python 2

  6. wdgwars-api-tester wdgwars-api-tester Public

    Systematic probe of the WDGoWars HTTP API surface. Detects outages, distinguishes route-not-bound from auth-rejected, and fingerprints styled 404 pages. Stdlib-only Python 3, single file.

    Python 1

Repositories

Showing 8 of 8 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…