How it's built

The Technology

The map-first architecture, the nine-agent AI swarm, the data model, and the Python brain beside the realtime spine.

06 — CORE APP ARCHITECTURE

Map-first, offline-first, private by design

Built on what Strivon already runs today (PWA + Supabase + on-device intelligence + Claude edge functions) and extended with a Python ML brain.

📱 Client (PWA / Capacitor native) Supabase — Auth · Postgres+PostGIS · Realtime · Storage Edge Functions (Deno) — coach · missions 🧠 Flask ML service — twin · embeddings · agents
Client

Vanilla-JS PWA today → Capacitor for native GPS/Health/notifications. Map-centric UI, offline journey capture, on-device GPS smoothing & sports-science math.

Data & realtime

Supabase Postgres + PostGIS for geospatial, Storage for media, RLS for privacy, Realtime for the anonymous network.

Intelligence

Deno edge functions for low-latency LLM calls; a Python/Flask service for the digital twin, vector memory (pgvector), and the agent swarm.

07 — AI ARCHITECTURE · THE AGENT SWARM

Nine agents, one companion

AgentJobInputsEngine
CoachEmpathetic daily guidance & check-insPlan, runs, mood, calendar, weatherLLM (Claude) + RAG
Memory / RAGRetrieve your past — “what was I doing last spring?”Journeys, notes, media captionspgvector embeddings
Mission GeneratorDaily adventures tuned to contextLocation, weather, time, fitness, historyLLM + rules + POI data
Route GeneratorLoops by distance, mood, scenery, elevationPostGIS, OSM, elevation, preferencesGraph routing + LLM
Behavioural AnalystDetect habit cycles, break points, personalityTime-series of activity & moodClassical ML (sklearn)
Wellbeing Insights“Which places lift you” correlationsMood deltas × place × routeStats / causal-lite
Digital TwinPredict future fitness, mood, burnout; simulate scenariosFull history + goalsTime-series forecasting
Story GeneratorNarrate your year / journey as a filmJourneys, memories, milestonesLLM + media pipeline
SerendipityAnonymous crossings & kindred spiritsHashed geohash grid, k-anonymityPrivacy-preserving match
The learning loop: every agent writes back to a per-user profile of motivations, triggers and preferred adventures. The companion gets more “you” every week — the compounding data moat.
← ExperienceBusiness →