A sports-technology venture set out to give everyday boxers the kind of frame-by-frame technical feedback that was previously available only to athletes with a full-time coaching team. The product needed to capture real training footage on a phone, understand what the fighter was actually doing — punches, guard, footwork, head movement — and return specific, data-backed coaching, all in a fast, native mobile experience.
The team built CornerCoach, a mobile-first, AI-enabled boxing training platform that pairs on-device computer vision with large language model reasoning. Boxers upload training videos, the app extracts technique metrics directly on the device, and those metrics are interpreted into structured, rated coaching feedback. At the core of the platform's language intelligence is Claude (Claude Opus 4.8), accessed via the Claude API.
Powered by Claude (Anthropic)
At the heart of the platform's coaching intelligence is Claude Opus 4.8, integrated through the Claude API. Where traditional fitness tooling relied on rigid thresholds and generic tips, Claude interprets the computer-vision output the way an experienced boxing coach would — reading the numbers in context, weighting what matters for the training type, and explaining why with reference to specific moments in the round.
The on-device vision pipeline produces a rich, structured metrics payload (punch counts and types, punches-per-minute, combination sequences, guard-up percentage and drop timing, head-movement activity, power/rotation estimates, footwork, and detected patterns). That payload is paired with Claude — fed into the model alongside a master coaching prompt and rating rubrics — and Claude returns decision-ready coaching.
Pose, tracking, punch detection & structured metrics payload
Master coaching prompt + rating rubrics + contextual reasoning
Rated feedback, drill recommendations & tappable timestamps
Interprets the CV metrics for each session and returns structured, rated feedback ("excellent / good / needs work") per category, with priority flags and a recommended drill.
References exact timestamps and numbers ("at 0:45 your guard dropped after the cross; cross rotation was 34°, aim for 45°+") so feedback is concrete, not generic.
Adapts its analysis to shadow boxing, bag work, pad work, or sparring, applying the right priorities for each (e.g. defense weighted highest in sparring; pad-contact accuracy treated as the key metric in padwork).
In sparring, reasons over both fighters' data to surface who controlled exchanges and the tactical advantages and vulnerabilities.
A separate Claude (Claude Haiku 4.5) integration analyzes production errors from the live monitoring pipeline, turning stack traces into plain-language root-cause summaries and suggested fixes for the engineering team.
Faster, more consistent technique interpretation than manual review
Coach-quality, moment-specific feedback available on every session
Reliable per-fighter analysis in sparring scenarios
Cross-session metrics captured in a normalized database for progress tracking
The team designed and built a native iOS app backed by a TypeScript REST API that integrates Claude-driven coaching with an on-device computer-vision pipeline and structured metrics storage. The system supports:
.webp){kind=icon}
Swift / SwiftUI (iOS 15+)
Express.js + TypeScript
Bun Runtime
Claude API (Opus 4.8 + Haiku 4.5)
MediaPipe Pose Landmarker
ByteTrack
1D-CNN Punch Classifier
PostgreSQL + Prisma
Google Cloud Storage
Apple StoreKit (IAP)
Map the coaching workflow, training types, and the metrics that actually drive boxing feedback.
Native iOS UX with an analysis-first flow, overlay viewer, and coaching results architecture.
On-device CV pipeline (pose → tracking → punch classification → metrics), Claude API integration for coaching, and normalized metrics storage.
Quality testing, credit/billing automation, monitoring pipeline, and cloud deployment.
Discovery & UX strategy
On-device CV pipeline & AI model integration
Testing, optimization & deployment
Agile methodology with iterative development and feedback — regular sprints, stand-ups, and progress tracking — with monitoring and AI-assisted error triage feeding directly back into the build loop.
From computer vision on mobile to Claude-driven coaching intelligence, we help sports and consumer platforms deliver expert-level feedback at scale.
Trusted by global partners
