Brandon
Stevison

A film and TV watchlist + social platform — a Letterboxd alternative with full series/season tracking, event-sourced watch history, threaded reviews, and region-aware streaming availability computed against each viewer's country and subscribed services.

Built solo, first commit to production in six weeks. Python/FastAPI + async SQLAlchemy + PostgreSQL, React frontend, Capacitor Android app, Docker/Caddy deployment with tiered automated backups. 75+ REST endpoints, ~20 zero-data-loss schema migrations against a live database.

The best stories came after launch: users importing 2,000+ film libraries exposed a first-traffic incident (retry storms, connection-pool starvation) diagnosed and fixed in one pass; a ratings-provider circuit breaker that treated "unknown film" as an outage and silently poisoned nightly refreshes; and a fixture suite that passed while real-world imports failed — because real Letterboxd exports don't look like synthetic ones.

A cross-platform arcade puzzle game shipped solo through seven production releases — eight game modes, real-time global leaderboards, anonymous-first auth, rewarded ads + IAP monetization, and the full store presence (listings, screenshot sets, ad creatives, a playable ad).

The analytics work is the point. A deliberately small custom event taxonomy — designed backwards from the questions a paid-UA campaign needs answered — feeds a Firebase → BigQuery export for SQL-level cohort, retention, and funnel analysis. Rounds are timed from first tap, not screen-open, so engagement numbers measure play, not navigation.

Review-prompt timing was iteratively tuned from player behavior: the naive v1 trigger asked new players for a rating mid-onboarding; the fix gates asks to celebratory moments with per-mode-family thresholds, behind a soft-ask layer that makes the funnel measurable despite the deliberately opaque native review API. Paid UA on Google Ads surfaced ~80% install concentration in South Asia at low CPI — which reshaped the follow-up Tier-1 geo strategy.

A Plotly Dash application for exploring target-position uncertainty when two platforms observe the same object: azimuth/elevation/range measurements transformed to Cartesian with Jacobian covariance propagation, fused via Julier–Uhlmann covariance intersection with Mahalanobis-distance gating, and rendered as interactive uncertainty ellipsoids on a drag-to-position map.

Representative of the day-job work: quick-turn analysis of 100–1,000 Hz instrumentation streams to support same-day field-test decisions, and pre-event dashboards that let non-specialists interrogate an algorithm before committing range time to it. Shown here as a pure-methods demonstration with synthetic geometry.