

ZOSA
A production AI system on iOS, Android and web: multi-provider orchestration, hybrid memory retrieval and real-time voice, architected and built by one person.
Six model providers behind one capability-tiered interface. A three-layer memory system with dynamic budgets and a relevance floor. Two independent real-time voice pipelines behind an identical contract. An intelligence layer that runs entirely off the response path. 819 commits across 291 separate days, still shipped to most weeks, alongside a full-time director-level role.
- Role
- Architect, sole engineer, product owner
- Scale
- 120 endpoints, 6 providers, 14 models, 3 platforms
- Impact
- Live on the App Store, Google Play and web, with paying repeat subscribers
- Python 3.12
- FastAPI
- LLM orchestration
- Vector retrieval
- WebRTC realtime
- React Native
- Firestore
- AES-256-GCM
- 01Problem
Suicide has taken people out of my life more than once: friends, and friends of friends. Each time, nobody close to them knew how bad it had got. Software does not fix that and I would not claim it does. It can occupy the gap: one hour a week with a therapist, and nobody there for the other 167.
- 02Challenge
A model that forgets is not a companion. Continuity across months means solving retrieval, context-window management and cost control at once, on a consumer product where a single voice session can burn real money in real time.
- 03Approach
Treat the model as one component in a system, not as the system. Design the layers around it: what gets retrieved, what gets summarised, what never reaches it, and what happens when it is wrong.
- 04Technology
A 120-endpoint async FastAPI service where the entire intelligence layer runs as background tasks. Six providers behind capability tiers resolved from environment variables. Hybrid retrieval over Firestore native vector search with a FAISS fallback.
- 05Result
A live product on three platforms with paying repeat subscribers, still shipped to most weeks, on an AI architecture that came out of primary research.
