Case study · Our product

How we built DC Rally

Your Game. Your Coach. Your AI Advantage. AI-powered badminton intelligence: a parent records a match on one phone, and DC Rally reads every serve, point, shot and movement, then turns it into a match report, a coach’s view and answers in plain language.

Built and owned by: Decision Culture
Sector: Youth sport, academies
Platform: Databricks lakehouse
Status: Academy pilot starting Nov 2026

Visit dcrally.ai How it works

1 phoneon a tripod. No court installation, no extra hardware.

99.9%of in-rally moments with the player located, checked on a full match

98%of rallies found, none split in two (283 of 290 across 12 matches)

9 in 10points with the server and the winner read right (server 100%, winner 93% on hand-checked matches)

Measured on our test matches against hand-checked truth. A separate set of matches is sealed for the final test and has never been used for tuning, so the pilot numbers will be reported on matches the system has never seen. On held-out matches during development, the server was read right 91% of the time and the winner 85%.

The opportunity

Junior players are filmed all the time. Almost nobody analyses the footage.

Parents of competitive junior players record nearly every match on their phones. The videos sit in camera rolls. A coach has no time to watch hours of footage for every child, and professional match analysis needs fixed multi-camera systems that only top venues can afford.

We set out to give every family and academy the kind of match intelligence that elite players get, from the phone they already own. It had to work at a club court on a Saturday, recorded by a parent, with the result back within 24 hours.

A phone mounted sideways on a tall tripod behind the baseline, filming the whole badminton court — The only hardware: one phone on a tripod behind the court.

Why this is hard

The same problems any real-world AI system meets: imperfect inputs, noise, and conditions that change every time.

One camera, set low

A phone behind the court sees the near player clearly and the far player small, so depth has to be inferred, not measured.

A tiny, very fast object

The shuttle is a few pixels wide, blurs at speed and disappears against lights and walls.

Busy halls

Neighbouring courts, other shuttles, and coaches sitting right behind the players all look like players to a model.

Real-world video

Phone footage arrives with variable frame rates, HDR, camera drift and cuts. Every match is different.

From one phone video to a full match read

Six stages, run automatically after the parent presses Start analysis. The parent’s only technical step is tapping four points on the court.

ParentRecord and uploadPhone on a tripod, head height or above, behind the court.
ParentMark the courtFour taps on the front service line corners map the video to real court metres.
AISee the playComputer vision tracks both players’ bodies and the shuttle in every frame.
AIRead the matchHits, rallies, serves and points, checked against the rules of the score.
LakehouseStore and governRaw, cleaned and report tables, each family’s data kept apart.
AppReport and askMatch report, coach comparison and Ask the Coach answers.

Report card

21-1718-2121-15

54%points won on own serve

61%points won receiving

1.4 kmdistance covered

31longest rally, shots

Six corners. Where the player was when each shot was struck. The back-left corner stands out: 19 shots, the most pressure came there.

Ask the Coach: “How did he do on his own serve in game 2?”He won 8 of 19 points on serve in game 2 (42%), down from 58% in game 1. Most lost points ended at the net.

What families and coaches get

A match report a parent understands and a coach can act on

Match report: score by game, who served and who won every point, points won on serve and on receive, distance covered, rally lengths.
Court maps and rally replay: where each shot was played from, how the player moved and how quickly they got back to base.
Coach view: a coach invited by the family compares the players in their group over time. Numbers only: coaches never see the videos.
Ask the Coach: questions in plain language, answered from that child’s match data, never from guesswork.
Body at the shot In development the real frame at each lunge, overhead and ready position, with stance and reach measured, for the coach to judge.

A complete AI product, built end to end by Decision Culture

What we built, layer by layer.

ML / AI

Computer vision

Deep-learning models find both players’ bodies (17 joints each) and the shuttle in every frame, then follow the court if the camera drifts.

Pose tracking at 30 frames a second
Shuttle tracking that fills short gaps where it is hidden
A zoomed second pass for the far player

Match intelligence

Reading the game

A chain of models and rules turns tracks into badminton: each hit, each rally, who served, who won the point and what shot was played.

Score-aware decoding: every point must agree with the rules of the score
Shot type shown only when the model is sure
Court geometry in real metres from four taps

Data engineering

Databricks lakehouse

Each match moves through raw, cleaned and report tables with automated jobs and a monitor that flags anything unusual.

Unity Catalog governance and per-family row filters
Test and production fully separated
Promotion only after a same-results check

BI + full stack

Parent and coach apps

A web app turns the report tables into a match report for the family and a comparison view for the coach, on laptop and phone.

Match details and court marking in the app
Phone camera check before recording
Sign-in, with each family seeing only its own data

GenAI

Ask the Coach

Questions in plain language are turned into queries on that family’s own match data, so every answer can be traced to the numbers.

Answers grounded in governed tables
Released only after passing a fixed 30-question answer key
Each family can only query its own child

Trust

Governance and monitoring

Every match is checked automatically for signs that a reading went wrong, and the family sees a caveat instead of a confident wrong number.

Score cross-checked with what the parent enters
Warnings for camera, court and serve-side problems
A monitor across every processed match

DatabricksUnity CatalogDelta tablesDatabricks Jobs and AppsGenie (AI/BI)PyTorchPose estimationShuttle tracking (TrackNet family)Python

Built like a product that families will trust, not a demo

The same engineering discipline we bring to client work, applied to our own product.

Sealed test matches

A set of matches is locked away and never used for tuning. Final accuracy is measured only on them, so the numbers we report are the numbers a new family will see.

The bar is set before the result

Every experiment has its pass mark written down before it runs. A change that misses the bar is not shipped, however close it came.

Same results before any switch

A new version replaces the old one only after both run on the same matches and the differences are explained point by point.

Test and production apart

All work happens in a separate test environment. Production changes only by a planned promotion with a rollback path.

Ground truth by hand

Hundreds of points and shots were checked by hand against the video, so every accuracy figure has a real reference behind it.

Every decision written down

Each stage has a design document recording what was tried, what failed, and why the current approach was chosen.

650+tracked code changes

10design documents, one per major stage

12matches in the test bench

24 hpromised turnaround per match

The same capabilities, applied to your business

DC Rally turns messy raw signal into decisions people trust. That is the work we do for clients, whether the signal is video, machines or customers.

In DC Rally	In your business
Computer vision on phone video	Quality inspection on a production line, shelf and store audits, safety monitoring from existing cameras
Reading hits, rallies and points from raw tracks	Turning clickstream, sensor or transaction events into journeys, funnels and outcomes you can measure
Databricks lakehouse with test and production apart	One governed source of truth for marketing, sales and operations data, promoted safely
Ask the Coach	GenAI assistants that answer business questions from your own governed data, with every answer traceable
Per-family access and children’s privacy	Role-based access, consent and compliance designed in from day one
Sealed tests and pre-set bars	Models you can defend in front of a CFO, measured honestly before they go live

The same pipeline that reads a rally can read a production line or a sales funnel.

Privacy by design

Children’s data handled with care

Videos stay private to the family. Nothing is published or shared.
Each family sees only its own child, enforced in the data layer and tested.
Coaches see numbers only, never videos or pictures, and only when the family invites them.
No child’s image or name is used in marketing.
All data lives in one governed lakehouse, with test and production kept apart.

Intellectual property

Our own methods, owned by Decision Culture

DC Rally’s code, models and data pipeline belong to Decision Culture. Several parts of the architecture are being evaluated for a patent filing, including:

Court calibration from four taps on a phone video, with a check of the camera’s height and view before recording.
Identifying the server of each point from body, racket and shuttle evidence, with confidence-based acceptance.
Reading points so that they always agree with the rules of the score.

Starting with junior badminton, built to grow

The first customers are families of competitive junior players and the academies that train them: people who already film matches and want to improve faster. A subscription per player keeps it affordable; academies bring groups of players at once.

Now · Nov 2026