All projects
Self-directed project · Live

AZIMUTH · Launch Market Intelligence Platform

DatabricksdbtPython · SQLStreamlit
7,659orbital launches, 1957 to today
2,352historical flights in the simulator backtest
0 to 7phases, each with a build-log post
39questions in the Ask AZIMUTH test
AZIMUTH Market Overview: orbital launch attempts per year and share of attempts in 2025
Live app · updated daily
Open the app Opens azimuth-launch.streamlit.app in a new tab. Best on a desktop screen.

What AZIMUTH is

AZIMUTH is a business intelligence platform that models the global orbital launch market. It is provider-agnostic by design: instead of tracking one company, it models the whole market, so that any launch vehicle can be set as the focus and benchmarked against the field. The same underlying model serves competitive benchmarking, program planning, and plain market research.

The guiding question is a stakeholder's. If you are evaluating a new heavy-lift vehicle, how does it actually stand against the rest of the market on cadence, reliability, reuse, and mass to orbit? That question is what every part of the model is built to answer.

The views

On top of the model, the app has five views:

Market Overview

The current state of the launch market across providers, vehicles, and mission types.

The Maturation Race

How vehicles progress from first flight toward operational reliability, compared across programs.

Scenario Simulator

Adjust assumptions and watch a program's competitive position shift.

Vehicles

One page per rocket: its record, recent form against its lifetime, versions, and full launch history, with one click to simulate its future.

Ask AZIMUTH

Questions in plain language, in English or Spanish, answered from the same data as the rest of the app.

Where it stands

All eight phases, 0 to 7, are complete and the app is live. Here is what each phase delivered.

Phase 0 · Scope & data model
Grain, dimensions, outcome and censoring split, and scope decisions are locked and documented.
Done
Phase 1 · Production backfill
The full launch history from Launch Library 2, landed in the bronze layer and verified idempotent.
Done
Phase 2 · Silver modeling
Staging model in dbt, a curated vehicle-class seed, six data-quality tests, and CI, all green.
Done
Phase 3 · Gold layer
Five dimensions and fct_launches in a dedicated gold schema, under referential integrity tests, rebuilt by CI.
Done
Phase 4 · Analysis
Cadence, market share, and reliability by flight number, with Wilson intervals, a Kaplan-Meier survival model, and a per-vehicle benchmark, all in SQL and deployed through CI.
Done
Phase 5 · Data operations
Daily incremental refresh, completeness check against the source, freshness check, and a warning for vehicles new to the classification seed. Runs on DuckDB in GitHub Actions since November 2026.
Done
Phase 6 · The app
Market Overview, Maturation Race, Scenario Simulator and a Methodology page, live on Streamlit, with the simulator backtested against history.
Done
Phase 7 · Ask AZIMUTH
Plain-language questions answered from the app's data, with vehicle pages and shareable simulator scenarios. Scored by a 39-question test after every deploy.
Done

Lakehouse architecture

The platform will follow a medallion architecture on Databricks: raw launch data will land in bronze, get cleaned and conformed in silver, and be modeled into an analytics-ready star schema in gold with dbt, with data-quality tests guarding every layer boundary.

Since November 2026 the daily refresh runs the same dbt project on DuckDB inside GitHub Actions, because Databricks Free Edition stopped reliably starting compute for scheduled jobs. The layers, models and tests are unchanged, and the output was checked identical before the switch. The build log post Moving the daily refresh off Databricks explains why and how.

Source data Launch Library 2 · Python ingestion 🥉 Bronze raw, as-landed Delta tables 🥈 Silver cleaned & conformed dbt tests 🥇 Gold star schema · dbt dashboard-ready
Medallion flow: ingestion → bronze → silver → gold

The data model

The gold layer is dimensionally modeled around a single fact table, fct_launches. The grain is one row per launch attempt, keyed on the launch UUID. Five conformed dimensions hang off it: date, provider, vehicle, orbit, and launch site. The mission attributes landed on the fact itself as degenerate attributes rather than a dimension of their own. Every question the platform answers (cadence per provider, reliability by vehicle, reuse rates over time, mass to orbit by destination) resolves to measures on that grain sliced by these dimensions.

Two relationships in the source are genuinely many-to-many, program and configuration families, so they are parked for a later version rather than forced into v1.

dim_date date_key year · quarter · month day · decade is_weekend dim_provider provider_key · provider_name sector · country founding_year launch_count fct_launches date_key · provider_key · vehicle_key orbit_key · location_key launched_at · net_precision mission_name · orbit_name · pad_name status_id · status_name scope_class · is_outcome_known is_success dim_vehicle vehicle_key · vehicle_name vehicle_full_name · variant manufacturer · vehicle_class is_reusable · maiden_flight launch_count dim_launch_site location_key · location_name country · celestial_body coordinates · pad_count launch_count dim_orbit orbit_key · orbit_name orbit_class (incl. "Not published")
Star schema · one row per launch attempt in fct_launches · built and tested

Timeline

Phase 0 · Complete · Aug 2026
Scoping & data audit

Read the Launch Library 2 API closely, drafted the medallion schema, made the scope decisions in writing, and ran a first data-integrity check. Deliverable: a written scope document in the repo.

Phase 1 · Complete · Aug 2026
Production backfill

Pulled the full launch history from Launch Library 2: 76 pages, 7,598 records, throttled to the API's rate limit over roughly five hours. Built the loader to be resumable and idempotent, with a load audit table recording every run. Verified by running the same load twice: the second run inserted 0 rows and updated 0 rows.

Phase 2 · Complete · Aug 2026
Silver modeling

Re-ran the Phase 0 data checks against all 7,598 launches, which overturned the orbital-scope rule and produced a curated, human-reviewed vehicle classification seed (485 configurations). Built the dbt staging model, six data-quality tests, and CI on GitHub Actions, green on the first real run.

Phase 3 · Complete · Aug 2026
Gold layer

Built the dimensional model: five dimensions (date, provider, vehicle, orbit, launch site) and the launch fact table, in a dedicated gold schema, under referential integrity tests, rebuilt by CI on every push. Providers enriched with country and sector, vehicles with disambiguated full names, and the orbit coverage gap made visible through an explicit "Not published" member.

Phase 4 · Complete · Oct 2026
Analysis layer

Cadence, market share, and reliability by flight number, by configuration and by family. Every reliability figure carries a 95% Wilson interval, plus a Kaplan-Meier survival model validated against Python's lifelines and an era-matched benchmark per vehicle. CI now deploys to production schemas after all tests pass.

Phase 5 · Complete · Oct 2026
Data operations

A daily scheduled refresh that loads only launches changed since the last run, then rebuilds, tests and deploys the models. Every run checks completeness against the source's own count and data freshness, and names any new vehicle that needs classifying.

Phase 6 · Complete · Oct 2026
The app

Market Overview, The Maturation Race and the Scenario Simulator, plus a Methodology page, live on Streamlit. Neutral by default, with any provider selectable as the focus. The simulator is backtested against how real vehicles performed after five past dates.

Phase 7 · Complete · Oct 2026
Ask AZIMUTH

Questions in plain language, answered from the same data as the app, with follow-up questions, a report button for wrong answers, and a 39-question test that runs after every deploy. The same phase added a page for every vehicle and shareable simulator scenarios.

Nov 2026
Daily refresh moved to DuckDB

Databricks Free Edition stopped starting compute for the scheduled refresh. The same dbt project now runs daily on DuckDB in GitHub Actions, with raw data kept as parquet in the repository. Output verified identical to Databricks before switching.

Build log

One post per phase

  1. November 13, 2026

    Moving the daily refresh off Databricks

    Databricks Free Edition stopped starting compute for the scheduled refresh. The same pipeline now runs every day on DuckDB inside GitHub Actions, with identical results.

    Read the post
  2. October 2026

    Phases 5 to 7: Daily refresh, the app and Ask AZIMUTH

    The data now refreshes itself every day, the app is live, and you can ask it questions. How each part works and how it was tested.

    Read the post
  3. October 2026

    Phase 4: Analysis layer

    What counts as one vehicle, reliability with confidence intervals, two pre-launch failures, and a model that was tested and not used.

    Read the post
  4. August 2026

    Phase 3: Gold layer

    The star schema, a "Not published" row for launches with no orbit, six more Moon launches, and the same float bug again.

    Read the post
  5. August 2026

    Phase 2: Silver layer, and a scope change

    dbt, the first silver model, tests and CI. Also: 7% of launches have no orbit, which changed how orbital launches are defined.

    Read the post
  6. August 2026

    Phase 1: Bronze ingestion

    All 7,598 launches from the Launch Library 2 API into a Delta table, and the problems along the way.

    Read the post
  7. August 2026

    Phase 0: Scope

    What the Launch Library 2 data can support, and the decisions written down before building anything.

    Read the post