Six device formats, one player, no unified record
Every golfer with a launch monitor has data trapped in a vendor silo. Garmin, Rapsodo, TrackMan, SkyTrak. Each exports a different CSV with a different schema, a different idea of what a "shot" is, a different way of naming clubs. The player's actual ability lived in the union of all of those, but no system could read the union.
We built parsers per vendor, a normalized shot model that treats device-source as metadata rather than as a wall, and a write-once-read-anywhere data layer that lets every downstream feature draw from one truth. Adding a seventh format becomes a parser, not a project.
Where this transfers: any business with multiple data vendors that don't reconcile. Sales tools, inventory systems, compliance feeds, claims processors.
A 36,000-course catalog with no personal context
The licensed course catalog provides industry-standard course geometry. It is expensive, generic, and identical for every customer who licenses it. Geometry alone tells a player nothing about how to play. It just tells them where the hazards are.
We built a strategy model that pairs that licensed geometry with the player's own dispersion patterns and per-club tendencies. The result is hazard guidance no other product on the market can produce, because no other product has both the substrate and the personalization layer. The licensed data became valuable only when fused with what was ours.
Where this transfers: licensed market data fused with internal forecasts. Industry benchmarks fused with customer-specific behavior. Any place a generic dataset becomes valuable only when paired with what's yours.
Paper scorecards in a digital system
Every golfer still uses paper scorecards. Manually entering them post-round is friction. Skipping them creates data gaps. The product had two choices: pretend the paper doesn't exist, or build a path through it.
We built the path. A phone photo runs through Claude Opus vision extraction, the result is reconciled in real time against shot-tracker data already in the database, and discrepancies are surfaced rather than hidden. Sub-two-second turnaround, paper to structured record.
Where this transfers: invoice OCR reconciled against purchase orders. Inspection photos reconciled against work orders. Anywhere paper or images meet structured systems.
One database, two questions, two answers
A golfer asking "what should I hit right now" needs different reasoning than the same golfer asking "what should I practice this week." The first question wants current form, recent misses, what's working today. The second wants trends, regressions, what to invest in. Same player, same data, two distinct windows.
We built two AI personas over one shot database. The Caddie reasons over a 30-day window. The Coach reasons over 90. One source of truth, two lenses, two recommendations that don't contradict each other because they're answering different questions.
Where this transfers: tactical dashboards and strategic reviews drawing from one warehouse. Sales coaching and territory planning from one CRM. Anywhere the same data needs to answer different questions on different time horizons.
Recommendations no one opens are recommendations that don't exist
Every analytics product is haunted by the dashboard no one opens. Insight without surface area is dead weight. The hardest part of this engagement was not the model layer or the data layer. It was making sure the system met the player in the moment, on the device in their pocket, with one tap.
We built native iOS and Android from a single React Native codebase, with contextual surfaces. Round Prep before the round. Caddie during. Coach after. The system shows up where decisions actually happen, not on a desktop hours later when the moment has passed.
Where this transfers: any team where the difference between "the data exists" and "the data is used" is whether the tool gets opened at 2 PM on a Tuesday.
