Live dashboard →  |  Source on GitHub →

A self-hosted analytics pipeline purpose-built for the Garmin Approach R10 launch monitor and the Garmin Golf app. Drop your range-session CSVs into Data/, run one command, and get a multi-page GitHub Pages dashboard with dispersion maps, gapping analysis, club path infographics, environment context, and personalised practice recommendations — all derived from your own R10 data.

# 1. Install (requires Python 3.11+)
pip install -e .

# 2. Export a session CSV from the Garmin Golf app and drop it into Data/
#    (see "Data Workflow" below)

# 3. Build the dashboard
golf-build build --data-dir Data --output-dir output/site

# 4. Open the dashboard
open output/site/index.html

Push to GitHub and GitHub Actions automatically deploys the same output to GitHub Pages on every commit.

1. Open the Garmin Golf app on your phone.
2. Tap Range → select the session you want to export.
3. Tap the share / export icon (top-right corner).
4. Choose Export CSV and save the file.
5. Copy the .csv file into the Data/ directory at the root of this repo.
6. Run the build command (or push to GitHub to trigger the Actions workflow).

The pipeline accepts both export formats the Garmin Golf app produces:

Tip: Keep one CSV per session. The pipeline processes every .csv in Data/ and aggregates trends across all sessions automatically.

Plots carry distance and smash factor over time with forecast confidence bands. Each session is also shown as a sortable row with carry, smash, offline, and outlier-rate stats.

A single per-club number derived from four R10 signals:

The weights are calibrated from the population of your clubs so the score is comparable across sessions regardless of R10 firmware updates.

Each club's drilldown panel includes a path/face relationship diagram — a visual showing your average club path direction, face-to-path angle, and the resulting ball-flight bias. A scatter chart plots every shot's club path against face-to-path so you can spot patterns and outliers at a glance.

Sorted carry distances across your bag are compared pair-by-pair:

• Bunching — clubs within 7 yards of each other, or whose carry SD windows overlap, are flagged.
• Distance void — a gap greater than 25 yards suggests a missing club or a loft that needs adjustment.

Carry targets are generated aspirationally from a club-type anchor ladder (Driver → woods → hybrids → irons → wedges) so the target always represents a realistic improvement goal rather than a mirror of your current average. A Progress To Goal slope chart shows every club moving from current carry to target carry in a single glance.

Each build captures temperature, humidity, air pressure, and air density from your session data and calculates:

• Carry vs. temperature slope — how many yards of carry you gain/lose per °F.
• Carry vs. air density slope — the density effect on ball flight across your sessions.

Rather than comparing your smash factor against a generic industry benchmark, the pipeline calculates your personal 90th-percentile smash factor for each club from your own historical shots. The recommendation engine only flags a strike-quality issue when your average falls more than 0.08 below your own best. The benchmark grows as your ball-striking improves.

A session-by-session bar chart showing whether your average lateral miss is trending left or right.

Ranked recommendations on the Coaching page, covering:

• Directional bias (avg_carry_deviation_distance)
• Clubface control (face_to_path_stddev)
• Strike quality (avg_smash_factor vs. your personal potential)
• Tempo stability (tempo_stddev)
• Outlier rate (anomaly model)
• Distance gapping (bunching and voids)
• Data quality (high flagged-shot rate)

Each recommendation includes a confidence score, confidence reason, and priority explanation so you always know why it's ranked where it is.

Per-session flagged-shot rates, missing critical-field counts, suspicious-reading breakdowns, and a doughnut chart of flag reasons — so you can trust the analytics driving your coaching.

Data/                          ← drop Garmin Golf app CSV exports here
src/golf/
  ingest.py                    ← CSV parsing, unit conversion, zero-shot cleaning
  analytics.py                 ← club/session summaries, targets, recommendations
  ml.py                        ← lightweight anomaly scoring
  site.py                      ← multi-page static dashboard generation
  cli.py                       ← `golf-build` entry point
  config/
    field_aliases.yaml         ← column-name aliases for all supported monitors
                                  (edit here to add new Garmin app column names)
tests/
  test_pipeline.py             ← pipeline integration + unit tests (52 tests)
.github/workflows/             ← GitHub Actions: build + deploy to Pages

The project uses uv and hatchling.

# Install uv (once)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create a virtual environment and install all dependencies
uv sync

# Run the build
uv run golf-build build

# Run tests
uv run pytest tests/ -v

# Add a new dependency
uv add some-package

Lockfile (uv.lock) is committed so builds are fully reproducible.

Garmin Golf app updates sometimes rename or reformat column headers. No Python changes are needed — just add the new mapping to src/golf/config/field_aliases.yaml:

aliases:
  # ...existing entries...
  "New Column Name (units)": internal_field_name

Run pytest tests/ -v to verify the change doesn't break existing test data.

All recommendations are explainable and derived from your R10 data — no black-box models:

Q: My R10 export has extra rows at the top before the column headers — will it parse? A: Yes. The loader scans every row and identifies the header row by looking for known column names and Garmin-specific anchor keywords (Shot Number, Club, Date, etc.). Metadata preamble rows are skipped automatically.

Q: Some shots show 0 for ball speed in the app but aren't in my analytics — is that right? A: Yes. Zero ball-speed shots are phantom detections (the R10 detected your swing but didn't track the ball). The pipeline nulls them out and excludes them from averages to prevent skewing your stats.

Q: Can I use this with a different launch monitor? A: Yes. Trackman 4, Rapsodo MLM2, SkyTrak, and FlightScope/Mevo+ are all supported via the same field_aliases.yaml config.

# 1. Install (requires Python 3.11+)
pip install -e .

# 2. Export a session CSV from the Garmin Golf app and drop it into Data/
#    (see "Data Workflow" below)

# 3. Build the dashboard
golf-build build --data-dir Data --output-dir output/site

# 4. Open the dashboard
open output/site/index.html

Push to GitHub and GitHub Actions automatically deploys the same output to GitHub Pages on every commit.

1. Open the Garmin Golf app on your phone.
2. Tap Range → select the session you want to export.
3. Tap the share / export icon (top-right corner).
4. Choose Export CSV and save the file.
5. Copy the .csv file into the Data/ directory at the root of this repo.
6. Run the build command (or push to GitHub to trigger the Actions workflow).

The pipeline accepts both export formats the Garmin Golf app produces:

Tip: Keep one CSV per session. The pipeline processes every .csv in Data/ and aggregates trends across all sessions automatically.

Plots every shot's carry distance vs. lateral offline deviation on an interactive scatter chart. Each club gets its own colour and a toggle button so you can isolate or hide specific clubs. Outlier shots (flagged by the anomaly model) are visually differentiated.

A single per-club number derived from four R10 signals:

The weights are calibrated from the population of your clubs so the score is comparable across sessions and clubs regardless of R10 firmware updates.

Rather than comparing your smash factor against a generic industry benchmark, the pipeline calculates your personal 90th-percentile smash factor for each club from your own historical R10 shots. The recommendation engine only flags a strike-quality issue when your average falls more than 0.08 below your own best. This makes the benchmark grow as your ball-striking improves.

Sorted carry distances across your bag are compared pair-by-pair:

• Bunching — clubs within 7 yards of each other, or whose carry standard-deviation windows overlap, are flagged. This is a common R10 finding when a player carries two similarly-lofted wedges or has inconsistent contact with short irons.
• Distance void — a gap greater than 25 yards between adjacent clubs suggests a missing club or a loft that needs adjustment.

A session-by-session bar chart showing whether your average lateral miss is trending left or right over time. Red bars = right miss, blue bars = left miss.

The R10 occasionally saves a phantom shot (swing detected, ball not tracked) with ball_speed = 0. The pipeline automatically nulls these values, excludes them from averages, and flags the session if more than 10% of shots are affected — prompting you to check sensor placement and lighting.

Up to eight ranked recommendations per build, covering:

• Directional bias (avg_carry_deviation_distance)
• Clubface control (face_to_path_stddev)
• Strike quality (avg_smash_factor vs. your personal potential)
• Tempo stability (tempo_stddev)
• Outlier rate (anomaly model)
• Distance gapping (bunching and voids)
• Data quality (high flagged-shot rate)

Data/                          ← drop Garmin Golf app CSV exports here
src/golf/
  ingest.py                    ← CSV parsing, unit conversion, zero-shot cleaning
  analytics.py                 ← club/session summaries and recommendations
  ml.py                        ← lightweight anomaly scoring
  site.py                      ← static dashboard generation
  cli.py                       ← `golf-build` entry point
  config/
    field_aliases.yaml         ← column-name aliases for all supported monitors
                                  (edit here to add new Garmin app column names)
tests/
  test_pipeline.py             ← pipeline integration + unit tests
.github/workflows/             ← GitHub Actions: build + deploy to Pages

The project uses uv and hatchling (migrated from setuptools).

# Install uv (once)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Create a virtual environment and install all dependencies
uv sync

# Run the build
uv run golf-build build

# Run tests
uv run pytest tests/ -v

# Add a new dependency
uv add some-package

# Add a dev-only dependency
uv add --dev some-dev-package

Lockfile (uv.lock) is committed so builds are fully reproducible.

Garmin Golf app updates sometimes rename or reformat column headers. No Python changes are needed — just add the new mapping to src/golf/config/field_aliases.yaml under the Garmin section:

aliases:
  # ...existing entries...
  "New Column Name (units)": internal_field_name

Run pytest tests/ -v to verify the change doesn't break existing test data.

All recommendations are explainable and derived from your R10 data — no black-box models:

Q: My R10 export has extra rows at the top before the column headers — will it parse? A: Yes. The loader scans every row and identifies the header row by looking for known column names and Garmin-specific anchor keywords (Shot Number, Club, Date, etc.). Metadata preamble rows are skipped automatically.

Q: Some shots show 0 for ball speed in the app but aren't in my analytics — is that right? A: Yes. Zero ball-speed shots are phantom detections (the R10 detected your swing but didn't track the ball). The pipeline nulls them out and excludes them from averages to prevent skewing your stats.

Q: Can I use this with a different launch monitor? A: Yes. Trackman 4, Rapsodo MLM2, SkyTrak, and FlightScope/Mevo+ are all supported via the same field_aliases.yaml config.