SplitStep · a product case study by Vibhav Gupta

A phone at the fence clears its throat and says:
"Long by two feet. 78 miles per hour."

SplitStep turns the smartphones players already own into a tennis coach, counter, and line judge — no hardware to buy, no cloud bill to pay, nothing to install on the court. This is the case for the idea: the market gap it attacks, the inventions that make it possible, and why the economics are structurally different from everything else in the category.

$0hardware cost to the player

$0cloud inference cost, forever

2phones = a sensing network

1patent application drafted

27→65%real-world ball detection, own models

11days, idea → court-tested

The gap in the market

Tennis sensing today makes players choose between price tiers, not product tiers:

tier	examples	the catch
Stadium systems	Hawk-Eye, installed smart courts	$10k+ per court, someone else's court
Court hardware	net-post cameras, sensor arrays	another gadget to buy, charge, mount, and lose
Single-camera apps	vision-only phone apps	one camera guessing at physics it can't see; fails exactly when light, angle, or a net-clipped ball defeats it
SplitStep	the phones already in your bag	—

Every incumbent treats sensing as a hardware problem. The founding observation: players already carry the hardware. A modern phone has a millisecond-accurate microphone, a fast camera, and a radio. Two of them, coordinated, are a small officiating system. The missing product was never a device — it was the orchestration layer.

The inventions

Insight 01 Two phones that only believe what they both heard

Any single microphone can be fooled — wind, pocket rustle, the next court over. The creative leap: make agreement between devices the filter. A sound only counts when two phones, placed apart, both witness it in a physically consistent way. Noise doesn't get filtered out; it never gets in. This idea — phones as corroborating witnesses — is the core of the drafted patent application, where the prior-art search found open white space.

Insight 02 Sound knows when. Vision knows where. The court knows how big.

Cameras sample time too coarsely to catch a bounce; microphones catch it to the millisecond but can't see. Fusing them on one clock measures what neither sensor can alone — and the court's own painted lines, with their regulation dimensions, turn four taps on a screen into full metric geometry. No tripods, no measuring, no calibration hardware: the court calibrates itself.

Insight 03 The second phone is the product, three times over

The same "spare" phone serves as: the strike microphone that unlocks serve speed and net-fault detection; the announcer in your pocket so calls are spoken where you stand, not across the court; and the remote viewfinder — a live view of the mounted phone in your hand, so you aim, zoom, and calibrate a camera whose screen you can't see. Competitors ship a companion app for score display. SplitStep makes the second device a sensor, a speaker, and a control surface.

Insight 04 Follow the courts, not the roadmap

The original wedge was wall practice — until real-world testing surfaced a market fact: practice walls are disappearing from Los Angeles parks. The product pivoted within a day to serve practice: solo-testable (no partner needed), playable on any court, immediately useful — and every serve recorded is training data that improves the vision models. A feature that funds its own R&D.

Insight 05 Accessories yes, gadgets no

A disciplined line on hardware: SplitStep will recommend a $15 fence clamp or the Bluetooth earbuds you already own (voice control from the baseline) — but never manufactures electronics. That keeps margins software-shaped, avoids inventory and support, and — as the freedom-to-operate research showed — steers around active patents that cover exactly the net-post camera gadgets others build.

The product today

Serve Practiceprop a phone behind the box, serve from the far end: spoken calls with distances, serve speed, net faults, and a placement map — hands-free with voice control

Wall & Rally Modeshit counting, streaks, personal bests, spoken milestones — solo or with two phones cross-checking

Own vision modelsa 1.16 MB ball detector trained from scratch on licensed + synthetic data; one training run ships to Android and iPhone identically

Works anywherephones pair directly even with zero connectivity — a court in a dead zone works exactly like one at a club

Strategic judgment

Patent position, built before launch

An adversarial prior-art campaign (commercial, academic, patent databases) mapped the landscape before a single public disclosure: the multi-device acoustic mechanism sits in identified white space, a provisional application is drafted, and two competitors' patent families were claim-charted to confirm freedom to operate.

A cost structure incumbents can't follow down

Everything runs on the player's own devices: no cloud inference, no per-user compute cost, full privacy by construction. Marginal cost of an additional user is effectively zero — and the product works offline, which cloud-dependent competitors structurally cannot match.

A data flywheel that respects the user

The models launch on licensed and synthetic data — no user footage required. But every practice session can (opt-in) contribute exactly the real-world imagery and acoustics the models lack. Usage compounds into accuracy; accuracy compounds into usage.

Execution leverage, managed like a team

I ran an AI engineering pair the way I'd run a high-output team: clear constraints, ruthless priorities, review gates before every release, and a standing rule that nothing ships without a number or a field test. Result: 29 shipped builds, 5 model generations, a patent draft, and a competitive teardown in 11 days — from one person's direction.

How I operate — mapped to the work I want to do

Product managementFound the wedge by testing on real courts, not in a deck. Killed my own favorite feature on market evidence ("practice walls are disappearing") and repointed the product in a day. Set one measurable gate for "it works" and refused vanity metrics — including re-measuring a competitor baseline from a claimed 83% down to its true 31–37% before trusting it.

Engineering managementPriorities ranked by user pain; adversarial review gates that caught release-killing bugs before the field did; same-day turnaround from field report to root cause to shipped fix; scope discipline — deferred features named explicitly with the evidence needed to unlock them. Managed IP risk like an engineering constraint: patent search and freedom-to-operate before public disclosure.

Data & ML engineeringBuilt the data supply chain end-to-end: licensed + synthetic training data (zero user-data dependency at launch), leakage-safe held-out evaluation, one attributable change per model generation (27%→65% real-world detection), operating-point tuning that halved false alarms without retraining, and a deployment pipeline that ships one set of weights to two platforms with verified numerical parity.

The vision

Serve practice is the wedge. The arc: two phones at opposite fences as a synchronized officiating pair — line calls with honest confidence bands, serve speed, match analytics, and a coaching layer on top — consumer Hawk-Eye, built from hardware the players already own. Feasibility work puts the core capabilities within a 12–18 month horizon, and the mechanism that makes it possible is the one already drafted for patent.

Where I fit

I'm looking for roles where this way of working compounds: product management, engineering management, or data engineering — especially where a product needs someone who can hold the market story, the technical reality, and the execution cadence in one head. SplitStep is what that looks like when I'm pointed at a problem I love.