The Domain Knowledge Gap
Without a background in cricket, I lacked the technical expertise to evaluate proper form, catch subtle biomechanical flaws, or know what specific corrections to guide.
0-to-1 AI Product Leadership
Building CricTrainer AI: From Monolith to Serverless AI Platform
How a CPTO reduced cloud burn by 98% while establishing AI agent guardrails.
$40/mo
98% Baseline Burn Cut
1 CPTO
5-Dev Agency Velocity
Closed Beta
1 Partner Academy
NVIDIA
Inception Program Member
- Company & Concept
- A flagship case study demonstrating 0-to-1 product strategy, agentic engineering, and GTM design. Built under WildWolf.ai to showcase production-grade computer vision, serverless cloud architecture, and MLOps capability.
- The Problem
- Out-of-class youth practice relies on unscalable, informal WhatsApp clips, leaving parents to manage home drills without structured, objective validation to keep young athletes engaged.
- The Leadership Story
- Founder/CPTO guiding CricTrainer AI from vision to production—self-funded path from agency pitch POC to solo build, replacing high-burn EC2 with pay-as-you-go serverless.
- The AI/ML Innovation
- Youth- and equipment-aware computer vision pipeline with strict agentic governance over coding tools and variable cloud COGS controls.
- Current Status
- Closed beta with an active partner academy, validating automated pose analysis and coach workflows across 6 core batting techniques.
01 // Hypothesis & Discovery
As a parent trying to help my son practice at home after youth cricket sessions, I experienced two major friction points firsthand—and quickly realized every parent on the sidelines was dealing with the exact same struggle:
The Parent-Child Practice Dynamic
Talking to other cricket parents—even those who played at high levels—revealed a universal truth: kids naturally resist instruction from their parents. Home drills frequently turned into arguments, making practice feel like a chore rather than an engaging routine.
Evaluating existing sports coaching apps revealed critical product flaws for the youth demographic:
Ad-Riddled UX
Existing consumer apps were heavily monetized with intrusive ads, creating a noisy, distracting environment unsuited for young athletes.
The "Adult Maturity" Assumption
Existing tools assumed adult self-motivation and discipline. They lacked mechanics to keep a 10-year-old engaged without direct parental enforcement.
The Pose Estimation Failure Point in Youth Sports
Testing standard off-the-shelf pose estimation models revealed severe limitations in youth sports. Models trained on adults failed due to distinct youth body proportions, while bulky protective gear (pads, helmets, gloves, bats) caused joint occlusion and broke standard skeletal tracking.
Privacy-by-Design Imperative
Realizing the primary end-user was a child meant that data security, video isolation, and strict COPPA compliance couldn't be afterthoughts or post-launch patches—they had to be foundational architectural constraints from Day 1.
Leveraging my background in mobile technology, gaming, and behavioral retention models, I framed this as a core product challenge: How do you make repetitive, technical practice accurate, objective, ad-free, and inherently fun for kids? I synthesized concepts across three emerging domains:
Video AI Frontier
While the tech industry was hyper-focused on text and static image LLMs in mid-2024, I recognized that real-time coaching required video understanding. I began exploring how computer vision and Vision-Language Models (VLMs) could grade temporal motion frame-by-frame, serving as an objective digital judge.
Equipment-Aware Biomechanical Analysis
Recognizing that standard off-the-shelf pose models break when dealing with youth proportions and equipment occlusion (pads, gloves, bats), I designed an architecture tailored for equipment-aware pose tracking and neutral, data-driven feedback to eliminate parent-child friction.
Behavioral Gamification
Adapting habit-forming retention loops (inspired by Duolingo) to transform dry biomechanical repetition into visual milestones, streaks, and engaging rewards tailored for youth athletes.
Core Product 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 Zero-to-One Launch Strategy (Mid-2024)
To translate this vision into a tangible product while managing full-time executive commitments, I authored a detailed PRD mapping out drill progressions, pose mechanics, COPPA privacy boundaries, and gamification loops. I brought in a specialized UX designer to establish a kid-friendly visual identity, then contracted a mobile development agency to assemble a rapid pitch-deck prototype. In mid-2024—before modern agentic coding tools had matured—this agency-led build served as a pragmatic baseline to validate user flows, and explore venture capital alignment.
02 // Commercial Realities, Strategic Pivot & System Evolution
Phase 1: Concept, Early Momentum & Strategy (Late 2024 – Mid 2025)
When I conceptualized CricTrainer AI under WildWolf.ai, getting accepted into the NVIDIA Inception program provided an encouraging early spark. It brought together key technical enablement resources and early validation that the idea had legs—making it worth putting initial capital into building a prototype.
At the time, I envisioned an ambitious consumer-heavy GTM model: a freemium mobile app for families, a standalone subscription tier for academies, and a multi-sided marketplace taking a commission on remote coaching transactions.
Phase 2: Commercial Realities & The Enterprise-to-Startup Mindset Shift (Q3 2025)
Evaluating early investor feedback and reviewing project burn brought critical commercial and technical realities to light:
Flawed Unit Economics & High Friction
Managing a 3-tier model (B2C, B2B, and a multi-sided marketplace) was overly complex, drove up Customer Acquisition Costs (CAC), and lacked clear unit economics.
The Infrastructure Wake-Up Call (Corporate vs. Startup Reality)
Coming from Fortune 500 corporate environments with enterprise cloud budgets, I initially selected a standard monolithic Node.js backend hosted on persistent AWS EC2 instances. Once I was self-funding, paying a $1,500–$2,000/month baseline server burn before onboarding active users was a sharp reality check. It forced an immediate shift from corporate tech norms to startup frugality, driving me to research and master event-driven, pay-as-you-go serverless architectures.
The Strategic GTM Pivot
Realizing that spending ad dollars to acquire B2C users on top of a fragile agency prototype was financially unsustainable, I killed the complex marketplace model and pivoted to a focused B2B2C strategy—selling directly to cricket academies as distribution hubs. Leveraging warm academy founder relationships as initial design partners allowed us to establish a repeatable B2B2C playbook before expanding outbound outreach, using academies to anchor the product and provide direct distribution to parents without high marketing spend.
Phase 3: The October 2025 Turning Point — From At-Home App to Coach's AI Assistant
In October 2025, when I demonstrated the prototype directly to academy coaches, I encountered a fundamental realization: coaches didn't want a generic, black-box AI telling their players how to bat.
Every academy and elite coach has a distinct technical philosophy—whether prioritizing classical English orthodoxy or modern power-hitting mechanics. A static, one-size-fits-all scoring engine created friction with coaches rather than collaboration.
This insight drove our final product pivot: transform CricTrainer AI into an AI Virtual Coaching Assistant powered by Coach Style Profiles. By allowing coaches to calibrate the AI to their personal biomechanical standards, the app became an extension of the academy itself—grading players against their specific coach's expectations and providing automated feedback in the coach's exact voice and style.
Technical Architecture & Product Shift: Agency Prototype vs. Production MLP
Agency Prototype· Feb 2025
- Product & Business StrategyGeneric B2C at-home practice app
- Cloud Infrastructure & EconomicsMonolithic Node.js on persistent AWS EC2 (~$2k/mo burn)
- Video Ingest StrategyHeavy server-side upload & cloud segmentation
- Biomechanical EvaluationRigid global baseline; hardcoded calculated score
- Feedback LayerNone (basic numeric output)
- Athletic NormalizationBroke on left-handed batters & equipment occlusion
- Coach AuthoringStatic / coaches excluded
Production MLP· 2026
- Product & Business StrategyB2B2C Virtual Coaching Assistant scaling individual coach styles
- Cloud Infrastructure & Economics100% AWS Serverless (Lambda, Step Functions, DynamoDB, Supabase) (~$40/mo burn)
- Video Ingest StrategyOn-Device VPM (ML Kit pose tracking on-phone for instant ready/finish clipping)
- Biomechanical EvaluationDynamic Coach Style Profiles & Calibration Matrix with cached JSON configurations
- Feedback LayerAmazon Bedrock (Nova Lite/Claude) generating feedback in coach's voice/style
- Athletic Normalization"Virtual Righty" Normalization (auto-mirrors LH coordinates into RH space)
- Coach AuthoringCoach Style Profiles Engine (dimension-targeted flawing & statistical compiler)
03 // Agent Governance & System Design
AI-Native Execution, Agentic Engineering & Governance
Replacing the agency prototype required engineering a multi-tenant platform spanning cross-platform mobile, coach admin portal, serverless cloud, LLM context engine, and RLS security. Delivering this scope as a solo builder required strict product governance over autonomous AI agents.
Building a complex cloud and computer vision platform as a solo builder required more than just using AI coding tools—it required strict product management governance over AI 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.
Governance Principle 01 // System Architecture
Planning Before Prompting
Leveraging my product leadership background, I used reasoning models (Gemini, Claude) primarily for architectural research, trade-off analysis, and system design before deploying autonomous coding agents (Cursor) to execute implementation. Clear PRD/TRD specs prevented agents from inventing unneeded complexity or drifting from the core architecture.
Governance Principle 02 // Human-in-the-Loop Control
Maintaining Builder Control & Engineering Rigor
Rather than blindly accepting generated code, I treated AI tools as junior engineers—rigorously reviewing diffs, enforcing modular design patterns, and retaining direct control over database schema design, authentication logic, and security policies.
Dynamic Coach Style Profiles
LLM context window loading of custom coach standards without straining database connection pools.
Spatial & Handedness Normalization
Vector coordinate transformation for left-handed athletes and equipment occlusion before scoring.
COPPA-Aligned RLS Isolation
Privacy-by-design data isolation at the database and storage layer for player data and video.
04 // Commercial Realities & Unit Economics
The shift from monolithic EC2 hosting to serverless cloud architecture eliminated fixed infrastructure waste, dropping our idle development operating cost to ~$40/month (down from a ~$2,000/month fixed server burn). However, building a scalable computer vision and LLM platform requires far more than celebrating a low idle server bill—it demands rigorous variable Cost of Goods Sold (COGS) modeling as usage expands.
Financial Realities: Fixed Baseline vs. Variable COGS Modeling
A single end-to-end cloud analysis pass—encompassing pose extraction, biomechanical scoring, and generative LLM feedback—ranges from $0.02 to $0.04 per video clip. Under unmitigated usage expansion, an active power user taking 1 daily practice session analyzing 6 basic shot types with 10 repetitions each (60 clips/day or ~1,800 clips/month) generates $36 to $72 per month in cloud processing costs, which would erode margins on a typical academy subscription baseline of $200/student/year (~$16.67/month).
Offloading Compute to the Edge
Moving pose tracking and frame segmentation on-device via local ML Kit reduces raw video payloads and eliminates cloud compute seconds spent on unpromising frames.
Efficient Prompt & Model Presets
Utilizing lightweight, optimized foundation model presets to minimize token overhead per coaching summary.
Smart Session Sampling & Usage Quotas
Establishing tier-based processing limits (e.g., analyzing the top 5 key shots per practice block) to deliver high-value coaching feedback while capping variable cost exposure.
COGS Breakdown
Raw power-user cost
$36–$72/mo
~1,800 clips/month at $0.02–$0.04 per end-to-end analysis
Post-mitigation target COGS
$1.50–$3.00/mo
Edge offload, model presets, and session sampling quotas
Target gross margin
70%+
On the $16.67/month academy subscription baseline
By offloading pose tracking to On-Device VPM and enforcing smart session sampling, target per-user COGS drops to ~$1.50–$3.00/month, unlocking a 70%+ gross margin on the $16.67/month subscription baseline.
Execution Velocity: Compressing Years into Months
Orchestrating agentic workflows under strict product governance enabled a single builder to achieve the output of a traditional 5-person engineering team. In less than six months, the platform evolved from a rigid agency prototype into a production-ready coaching engine—replacing static UI screens with real-time motion tracking and Coach Style Profiles that deliver personalized feedback.
Market Validation & Partner Alignment
Closed-Beta Pilot Scope
Deployed in closed beta with 1 anchor partner academy, onboarding 3 coaches and 25 youth players across 500+ analyzed drill sessions to validate automated pose analysis and Coach Style Profiles.
Coach Empowerment over Displacement
Rather than viewing the AI as an unwelcome competitor or generic grading tool, coaches embrace CricTrainer AI as an operational multiplier that scales their personal brand beyond physical practice sessions.
Academy Retention & Student Adherence
Academies gain a modern, tech-enabled retention tool that keeps young athletes practicing correctly at home.
Parental Trust & Friction Removal
Parents receive transparent, objective feedback that eliminates backyard friction while knowing their child's video data is locked down under child safety privacy controls.
Quantitative Outcomes & Leadership Impact Summary
| Metric | Prior Model | Production MLP | Impact |
|---|---|---|---|
| Monthly Fixed Cloud Burn | ~$2,000/mo | ~$40/mo | 98% reduction |
| Engineering Headcount | 5–6 agency | 1 CPTO + agents | ~80% lower capital cost |
| Time-to-Market | 9+ months est. | under 6 months | Faster pivot across 5 systems |
| Data Privacy | Manual / server-side | Automated RLS + edge | COPPA-aligned architecture |
| Unit Economic Modeling | Negative / unclear | Capped variable COGS | Protects academy unit economics |
05 // Personal Operating Principles
Building CricTrainer AI from zero-to-one provided practical validation on agentic execution, cloud unit economics, and product strategy. Beyond shipping a production-ready platform, the build refined four core operating principles that guide how I lead product and engineering organizations:
01 // Product Governance
Product Discipline Governs AI Velocity
AI development tools excel at execution speed, but velocity without clear direction is just accelerated technical debt. Senior product leadership—manifested through strict PRD/TRD specs, modular system design, and rigorous code reviews—is the single most important factor in keeping agentic tools aligned with core business goals.
02 // Unit Economics
Frugality Uncovers Granular Unit Economics
Transitioning from Fortune 500 corporate budgets to self-funding forced a deep audit of variable cloud compute costs. That constraint drove the early identification of power-user COGS risks, forcing architectural innovations like On-Device VPM and prompt optimization before scaling user acquisition.
03 // Customer Empathy
Customer Empathy Beats Technical Ego
The initial assumption that an AI app could unilaterally replace parental or coaching guidance fell flat when tested against market realities. True product-market fit was unlocked only when we humbled the original vision, listened to coaches on the ground, and built a system designed to amplify human expertise through Coach Style Profiles.
04 // Continuous Evolution
MLPs Are Baselines: Building Proprietary Moats
Modern AI products are dynamic feedback loops, not static releases. We leveraged third-party foundation models for 0-to-1 time-to-market speed, but every coach calibration and session review automatically builds a proprietary dataset. This enables Phase 2 transition to lightweight, custom-trained models—driving down API costs and securing a defensible IP moat over time.
Looking for a product leader who bridges board-level strategy with hands-on AI/ML engineering?
Get in touch on the executive profile →