25,000+ runners. $51K revenue. One platform that cheers, coaches, and scales.

Problem 1 — Emotional gap on race day: Friends could not be at the course, and existing apps did not stream live encouragement. Intellixa built shareable race codes, public spectator APIs (cheer, snapshot, SSE stream) without accounts, Redis-backed GPS fan-out, and a Bull-managed cheer queue so audio never stacks destructively during marathons.

Problem 2 — Generic coaching: Users needed plans that respect experience level, favorite distances, running device, units, and club context. Intellixa implemented a RAG pipeline: embed and retrieve relevant user and activity context, augment LLM prompts, and output structured exercise and nutrition recommendations that update as new races and preferences arrive—not static content.

Problem 3 — Scale and revenue: Live features and AI must run profitably. Intellixa delivered OTP + Firebase auth, subscription verification (Google Play, Apple, Superwall webhooks), cron hygiene for races and entitlements, and observability-ready Node.js services—supporting 25K+ users and $51K USD revenue on a stack built to grow.

CheerMy Run’s coaching layer uses retrieval-augmented generation—not a single prompt to a chatbot. At onboarding and over time, the platform collects structured runner context: demographics, experience, running device, shoe and gear preferences, distance goals, units (km/miles), audio preferences, and run-club affiliation. Race history, cheer engagement, and Strava-linked activity enrich the retrieval corpus.

When a user requests training guidance or diet planning, the system retrieves the most relevant slices of their profile and history from a vector index, composes an augmented prompt, and generates customized exercise blocks and meal guidance via LLM orchestration—with guardrails so outputs stay fitness-safe and on-brand. Plans evolve as new data arrives; the model does not hallucinate a generic marathon schedule for a 5K beginner.

Clients include a Flutter mobile app for runners and a lightweight spectator web experience via deep link. The API layer is Express (/v1) with Socket.IO per race, SSE for public streams, and Bull workers for cheers and async AI jobs.

Data plane: MongoDB for durable users, races, notifications, subscriptions, and recommendation artifacts; Redis for live telemetry, pub/sub, OTP rate limits, and queue backing; vector storage for RAG retrieval; AWS S3 for audio, TTS, and GeoJSON routes. AI plane: embedding models, vector search, LLM APIs, AWS Polly SSML, and ffmpeg for voice-merge pipelines.

During active races, Redis is source of truth for GPS; MongoDB syncs on schedule and on endRace. AI workloads stay off the hot path—recommendations generate asynchronously while live races stay sub-second.

When a runner updates location—REST progress or runner_location_update sockets—the backend validates the race is started, writes race_data:{code} in Redis, appends route segments with dynamic caps, and publishes to race:{code}:location. Spectators receive pace, distance, and map position without hammering MongoDB.

Cheer events broadcast on Redis timelines for spectator UI parity; FCM delivery follows queue spacing (default 10s per runner, 5s minimum gap, backpressure beyond five pending jobs). The result: marathon-scale cheer storms without stacked audio or push throttling meltdowns.

Beyond RAG coaching, CheerMy Run ships multiple AI layers for race day. AWS Polly (Joanna en-US, Mia es-MX) renders SSML-tuned cheers for outdoor audio; FCM payloads can trigger on-device TTS for instant text cheers. ffmpeg normalizes and concatenates multi-clip spectator voice notes into single MP3s on S3.

Post-race, RouteOptimizer denoises GPS, compresses polylines, and exports GeoJSON for sharing and Strava upload—clean maps and smaller payloads at scale.

Strava integration complements live cheering: OAuth with encrypted tokens, post-race activity upload with optimized polylines, and social proof in the athlete’s permanent feed. Zero-login spectator endpoints drove viral race sharing; personalized AI coaching drove return visits; Superwall plus native IAP converted engaged users.

Intellixa wired subscription state to MongoDB, daily expiry cron, and optional assertActiveSubscription gates—aligning product loops with verified revenue across 25,000+ users.