A software radio in a window in northwest Las Vegas listens to aircraft transponders on 1090 and 978 MHz. Its observations stream into BigQuery, where dbt turns them into tested, documented definitions: what counts as an aircraft, a sighting, "military", "identified". A public website, an MCP server for AI agents and a weekly AI-written report all read the same definitions, and data-quality checks email the operator when something breaks.
- Website: skyscanner.pharrhomelab.com
- MCP server:
https://mcp.skyscanner.pharrhomelab.com/mcp(Streamable HTTP; no sign-in except forrun_sql, which needs a token)
One definition per number. The core measures (a confirmed aircraft, a
sighting, military, identified, listening time) are defined once in dbt,
tested on every build and exported with their descriptions. The website's
Definitions page and the MCP server's get_definitions tool serve the same
text, so a person and an agent asking "how many aircraft were heard?" get
the same answer.
Data quality that alerts. dbt rebuilds hourly and checks source freshness every 30 minutes on Cloud Run. Failures become structured log lines, which log-based alert policies turn into email. Results land in BigQuery and drive the Data health page.
Permissions enforced by the warehouse. The website, the MCP server and the
agent each run as their own service account, and each can read only the
public dataset, made of authorized views (the agent also writes its
editions to ai). Private owners are never named, and aircraft on the FAA's
LADD list or using privacy addresses are flagged. A dbt test checks on every
build that no private owner's name reaches a public view, and emails an
alert if one does.
AI access with evals. The MCP server exposes 19 tools over the public views and serves the dbt definitions and column docs, so agents work from the same meanings as the website. Every Monday an agent (MiMo-V2.6-Pro through OpenRouter) investigates the week through those tools and writes the Notable aircraft page. A verifier checks every number, time and name against the tool output the agent cited and removes sentences that do not match. Evaluations on a fixed week measure Q&A accuracy against gold SQL (15/15) and coverage of a hand-written reference report (76%); results are on the Agent evals page.
| Notable aircraft, written by the agent | One aircraft and its sightings |
|---|---|
![]() |
![]() |
HOME (edge/) GOOGLE CLOUD (infra/terraform)
HackRF ─▶ receiver ─▶ ingest ─▶ Postgres ◀─ enrichment (FAA, tar1090-db, OpenSky,
│ plane-alert-db, Wikipedia)
replicator ── streaming inserts, ~30 s ──▶ BigQuery raw
│
Cloud Scheduler ─▶ Cloud Run job: dbt (hourly build, freshness every 30 min)
▼
analytics: staging → intermediate → marts
▼
public: authorized views (owners masked, LADD/PIA flagged)
│ │ ▲
Streamlit website ◀───┘ └──▶ MCP server ──┤
(Cloud Run) (Cloud Run) │
▲ │
Cloud Scheduler (Mon) ─▶ Cloud Run job: notable agent ── MCP ──┘ │
(OpenRouter) → verifier → ai.notable_editions ┘
dbt logs ─▶ log-based alert policies ─▶ email
| Directory | What it is |
|---|---|
| edge/ | The receiver, decoders, ingestion, enrichment, local Postgres and the BigQuery replicator, run with Docker Compose at home |
| dbt/ | Staging, intermediate, mart and public models; tests; MetricFlow semantic models and metrics; column docs generated from one table |
| jobs/dbt_runner/ | Runs dbt on Cloud Run, logs failures for alerting and stores run results |
| services/web/ | The Streamlit website |
| services/mcp/ | The MCP server (Streamable HTTP on Cloud Run, or stdio locally) |
| jobs/notable/ | The weekly agent, its verifier and its evals |
| infra/ | Bootstrap script (project, APIs, budget, Terraform state bucket), access-tier check, and Terraform for everything else in Google Cloud |
| docs/ | The runbook (alerts and what to do, deploying, recovery, cost guardrails, drill log) and the README's screenshots |
The Python components (edge/, services/web/, services/mcp/,
jobs/notable/) each have their own requirements.txt and tests
(python -m unittest discover -s tests). Linting is shared: ruff for
Python (pyproject.toml), sqlfluff for the dbt SQL (dbt/.sqlfluff) and
terraform fmt, pinned in requirements-lint.txt and run in CI with the
tests.
- Edge: see edge/README.md.
- dbt: install
dbt-core==1.12.6 dbt-bigquery==1.12.1(the versions injobs/dbt_runner/Dockerfile), then fromdbt/, with Google credentials:dbt depsanddbt build --profiles-dir .. The defaultdevtarget writes todev_analyticsanddev_public. - Website:
streamlit run app.pyfromservices/web/(PUBLIC_DATASET=dev_publicto read dev). - MCP server:
python server.pyfromservices/mcp/speaks stdio, for a local MCP client;MCP_TRANSPORT=httpserves Streamable HTTP. - Cloud:
infra/bootstrap.shonce, thenterraform applyfrominfra/terraform/.
The FAA aircraft registry (public domain), tar1090-db, the OpenSky aircraft database, plane-alert-db (ODbL), VRS standing data (CC0), War Watch Intel, ICAO Doc 8643 type designators via OpenSky, and type photos from Wikimedia Commons, credited on the site. Aircraft positions are received directly over the air.
The code is under the MIT License. Data sources keep their own licenses, listed above.


