Building CricTrainer

Living the problem

As a parent helping my son practice after youth cricket sessions, I hit two friction points every sideline parent recognizes: without cricket domain expertise I couldn't judge form or prescribe corrections—and kids resist instruction from parents, so home drills often became arguments instead of routines.

Market & competitive discovery

• Consumer coaching apps were ad-riddled and unsuitable for young athletes.
• Tools assumed adult self-motivation—not engagement mechanics for a 10-year-old.
• Off-the-shelf pose models trained on adults fail on youth proportions and equipment occlusion (pads, helmets, gloves, bats).
• Because the end user is a child, COPPA-aligned privacy had to be Day-1 architecture—not a post-launch patch.

Product & AI vision

Core challenge: make repetitive technical practice accurate, objective, ad-free, and fun. Synthesis across video AI / VLMs for temporal motion grading, equipment-aware biomechanics, and Duolingo-style gamification for youth retention.

Hypothesis

If we replace subjective parental oversight with an objective, gamified video AI engine—built on a privacy-first, ad-free architecture—we can eliminate home-practice friction, boost youth training adherence, and give coaches a reliable out-of-class validation tool.

Initial 0-to-1 launch (mid-2024)

While balancing full-time executive work, I authored a detailed PRD (drills, pose mechanics, COPPA boundaries, gamification), hired a UX designer, and contracted a mobile agency for a pitch prototype—validating flows, earning NVIDIA Inception acceptance, and exploring VC alignment before agentic coding tools matured.

Phase 1 — Concept & early momentum

NVIDIA Inception provided compute and momentum. Early GTM envisioned freemium B2C, academy subscriptions, and a multi-sided coaching marketplace—too complex for early unit economics.

Phase 2 — Enterprise-to-startup mindset shift

• 3-tier GTM drove CAC and unclear unit economics.
• Fortune-500 habits led to monolithic Node.js on persistent EC2—$1.5–2K/month burn before users—a sharp self-funding wake-up call.
• Killed marketplace complexity; pivoted to B2B2C via cricket academies as distribution hubs and design partners.

Phase 3 — From kids' app to coach's AI assistant

Coaches rejected a one-size-fits-all black-box scorer—each academy has a distinct technical philosophy. Final pivot: AI Virtual Coaching Assistant powered by Coach Style Profiles that grade in the coach's voice and standards.

Agency prototype vs production MLP

Scope: five systems, solo

Replacing the agency build meant a multi-tenant ecosystem: cross-platform mobile, coach admin portal (Next.js), serverless motion pipeline, context-aware LLM engine, and RLS-secured data architecture—work that traditionally needs a multi-disciplinary team.

Speed requires governance

Orchestrated Cursor, Claude, Gemini, and ChatGPT/o1 under an AI-native framework with managerial guardrails over agents.

AI coding tools are exceptional execution multipliers, but without product architecture discipline they can rapidly build the wrong things—or destroy the right ones.

Lessons in agent control

• Plan before prompting: reasoning models for architecture/trade-offs; coding agents only after PRD/TRD clarity.
• Hard lesson: an agent wiped a database during prototyping—backups recovered it; agents lost unmonitored write/drop privileges on stateful resources.
• Human-in-the-loop gates for schema/security changes; automated pre-migration snapshots.
• Treat AI as junior engineers—review diffs; retain control of schema, auth, and security.

Architecture highlights

• Coach Style Profiles & calibration matrix loaded into LLM context without straining DB pools.
• Spatial / handedness normalization for LH athletes and equipment occlusion.
• COPPA-aligned RLS isolation for player data and video.

Unit economics discipline

Idle burn fell to ~$40/month, but variable COGS matter: ~$0.02–$0.06+ per end-to-end clip. A power user (~1,200 clips/month) can generate $35–$70+ in processing vs ~$16.67/month at a $200/year academy seat—so margin protection became design criteria.

• Edge offload via on-device ML Kit for segmentation/pose.
• Lightweight model presets to cut token overhead.
• Session sampling & usage quotas (e.g. top shots per block).

Velocity & market validation

In under six months, one CPTO + agentic workflows replaced agency dependency and shipped a production coaching engine. B2B2C positioning made coaches partners—not displaced experts—while academies gained retention tooling and parents gained objective, private feedback.

Outcomes summary

Immediate product horizon

• Finish on-device VPM QA for shot segmentation and lower latency.
• Scale academy rosters with leaderboards and reward redemption.
• Extend Coach Style Profiles to additional sports/movement disciplines.

Executive takeaways

• Product discipline governs AI velocity—PRD/TRD, modular design, and code review keep agents aligned.
• Capital constraints drive superior architecture—self-funding forced resilient serverless design.
• Customer empathy beats technical ego—PMF arrived when AI amplified coaches instead of replacing them.