Building Pelaris
Full-Stack AI, Built from Scratch
I took the AI architecture patterns I'd built at enterprise scale and applied them to adaptive fitness coaching. One engineer, zero shortcuts.
The Problem I Saw
Most fitness apps are trackers pretending to be coaches. They count reps and log miles but never adapt. A triathlete's Tuesday run has nothing in common with a powerlifter's Tuesday squat, but the app treats them identically.
The gap between tracking and coaching is an AI problem. Context-aware reasoning over structured data with real-time adaptation. That's exactly what I'd been building at enterprise scale.
The Transfer
After Gartner keynotes, ServiceNow Knowledge presentations, and Forbes features, I kept seeing the same pattern: adaptive intelligence that deflects 94% of support tickets uses the same reasoning a good coach does.
Read the signals, adjust the plan, keep moving forward. The technology existed. I just pointed it at training.
AI-Driven Program Generation
The core engine is a 3-stage server-side generation pipeline built on Vertex AI (Gemini). Stage 1 builds a periodization strategy from the athlete's goals, sport, and methodology. Stage 2 generates week-by-week session plans. Stage 3 populates individual exercises with sets, reps, and coaching notes, all validated against methodology constraints before committing.
28 real training methodologies across 7 sports. 5/3/1, USRPT, Polarised Training, Sweet Spot, Coggan Zones. Each with its own distribution rules, deload protocols, and key constraints that the AI must respect. Not templates. Actual methodology-aware intelligence.
AI-generated week plan with per-session coaching notes, exercise selection, and methodology-aware periodization
Real-Time Tracking & Adaptation
The Flutter client handles live workout logging with set-by-set tracking, RPE capture, and exercise history. Every completed session feeds back into a Firestore-triggered analysis pipeline that compares actuals against targets.
RPE spikes for three weeks? The system adjusts intensity. Shoulder playing up? AI-powered exercise swaps from the methodology's approved alternatives. Missed sessions get recalculated, not guilt-tripped.
The feedback loop is server-side first. Cloud Functions triggers process completions, check-ins, and benchmarks without any client-side coordination.
Monthly AI Intelligence
Every month, a scheduled Cloud Function aggregates check-ins, session completions, benchmark PRs, and readiness data into a context snapshot. From that, it generates an AI-powered training summary with pattern recognition and forward-looking recommendations.
The context pipeline is the heart of the system. A TrainingContextSnapshot pulls methodology rules, duration constraints, RPE trajectories, sport benchmarks, and workout log insights into a single object that feeds every AI interaction.
All of this runs through a PII scrubbing pipeline before any AI processing. Privacy isn't a feature. It's architecture.
The Stack
Flutter for cross-platform (web-first). Firebase for real-time data and auth. Vertex AI (Gemini) for coaching intelligence. Cloud Functions for server-side generation, triggers, and scheduled jobs. CI/CD via GitHub Actions. Infrastructure as code.
One person. Designed, architected, and built from scratch. No team, no funding round.
By the Numbers
- 7 sports supported
- 28 real training methodologies
- 52 pre-built programs
- 3-stage AI generation pipeline
- Server-side first, triggers, not polling
- Privacy-first, PII scrubbed before AI
See it in action
Full-stack AI fitness coaching. No credit card required.