Cosmicly

I built Cosmicly as a private reflection journal: a person starts with a feeling, adds a life area and optional situation context, then receives a Finding they can save and return to. I owned that journey together with the clients and services behind it.

I used astrology as an interpretive timing lens for reflection and kept the experience human and legible. The system combines exact sky calculations with structured model output to give each check-in its context.

I made situation details optional so a person could choose how much context to share. I carried that control into the product architecture, with authenticated journal access and Row-Level Security around private data.

I built and tested the astronomy engine for planetary positions, charts, and timing calculations, then separated it from the models that generate interpretation. This was a product decision: exact results needed a stable source of truth as the AI experience evolved.

I gave calculations their own tests and model workflows their own contracts, validation, retries, and observability. That separation let me investigate calculation defects and generated-output failures independently.

I designed nine schema-constrained Gemini text and vision workflows around specific product jobs. Prompt and context engineering, guardrails, and orchestration gave each workflow a known response shape the Flutter clients could use.

I routed work across Gemini Flash-Lite, Flash, and Pro according to the workload, with structured-output validation, malformed-JSON recovery, bounded retries, timeouts, and region-aware fallback. Per-feature token and cost observability let me judge reliability and spend at the feature boundary.

When a long-form blueprint response was cut off, I traced the incomplete user result through the model path and found a workflow still carrying a 3,000-token limit. I corrected it to 8,192 so that workflow could return the full response.

I then centralized model configuration to stop individual features silently retaining inconsistent settings. The investigation became an architectural improvement as well as a fix for the immediate truncation defect.

I built the Flutter iOS and Android clients together with TypeScript/Deno APIs and a production Supabase/PostgreSQL backend. I owned authentication, Row-Level Security, realtime data, database functions and triggers, versioned migrations, and idempotent subscription webhooks.

I integrated RevenueCat billing, push notifications, analytics, crash reporting, and five-language localization, then built automated backend and client tests and staging/production CI/CD. These were parts of one product whose client, data, model, and subscription behavior had to work together.

I reached store-release readiness in roughly two to four months through 10+ builds and shipped on iOS and Android. The project extended my ownership from native clients into APIs, data integrity, model operations, and paid-product delivery.