Tyler Martin

First commit to public release in two weeks. Version 1.0 was cut and pushed into Google Play closed testing on day one, so the mandatory fourteen-day test window ran alongside development instead of after it. It has been in continuous live-ops ever since — 45+ player-facing versions across 64 store builds in under four months.

• Live service. Three-day rotating events with per-event leaderboards, daily and weekly quest resets on fixed timezone boundaries, cloud saves and online leaderboards on Unity Gaming Services, Play Games authentication, push notifications.
• Save architecture. Three local layers — PlayerPrefs, an atomic JSON write via temp/fsync/rename, and a backup — with forward-migrating schema versioning and a cloud-push guard so a blank save can never overwrite real progress. I wrote it after tracing an actual player data-loss incident through production logs and recovering the account.
• Economy under load. Found and fixed a bug where 64-bit integer overflow produced negative balances and false "affordable" states, then converted the entire economy to double precision end to end — wallet, UI, persistence — without breaking a single existing save.
• Monetization that survived an outage. When the ad network suspended the account, I moved monetization behind a provider abstraction with runtime fallback rather than shipping a panic update. It kept earning through the outage.
• iOS without a Mac. Built the App Store pipeline on Unity Build Automation and Codemagic, on hardware I don't own. iOS shipped from 1.7.0 and has taken 17 builds since.

Porting other people's mods

I port large third-party mods between loaders — the Mystical Agriculture stack, and Occultism at the original author's public invitation. That means reading an unfamiliar codebase end to end, mapping loader-specific registry, networking, capability and mixin APIs onto their equivalents, and preserving save and recipe compatibility so existing worlds survive the switch.

To pick targets I wrote a zero-dependency Node scraper against the GitHub API that finds permissively licensed mods with unmet demand and verifies the license before I spend a week on something I can't legally publish.

Contributions to other teams

Credited team member on nine third-party projects with more than half a million combined downloads — including The Broken Script (440K+), Spectrum_11 (42K+) and Naruto X Boruto (11K+) — working as programmer, commissioned developer, or content designer depending on what the project needed.

I also authored the quest books for Ancient Kingdoms 3 and 4 and KN: Mythic Classic: large-scale progression and content design built inside someone else's systems, on someone else's release schedule.

This is where most of my collaborative work lives. Contributing to another team's mod or pack means reading their conventions before writing a line, matching a codebase I didn't design, taking review from people whose project it is, and hitting a release date I don't control. It's the closest thing to studio work I've done, and it's the part I'd bring on day one.

You pick a class in the lobby and become whatever the dungeon gives you — the build is what dropped and what you took at level-ups. Floors escalate on a heat timer that eventually stops being survivable, so the run ends when the party decides to leave. Proximity voice is a design pillar rather than a feature: the party is the content, and you can hear each other from two rooms away.

The performance work is what I'd point a studio at. Turning graphics down should cost detail, never turn the game into flat untextured shapes — a weak machine should keep the look and buy back frames:

• Many fine knobs, not one Low/Medium/High switch. Presets only prefill fields; every setting stays individually adjustable and editable live from the dev console. The order things get cut in is the actual design: render scale with an upscaler first (the best frames-per-visual-loss there is), then shadow distance, then occlusion, then texture memory. Never "is it textured at all".
• A light budget, because the measurement said so. A floor carries ~90 point lights and profiling found 60 inside the camera frustum against a budget of 24 on the target GPU — over budget before a single shadow existed. Torches now sort by distance each tick: the nearest few light, fewer still cast shadows, the rest switch off. Culling uses hysteresis (on at the cull distance, off at 15% past it) so standing on a boundary doesn't strobe, and unlit torches skip their flicker update, taking ~90 Update calls per frame down to the handful you can actually see.
• Voice chat written from the spec. IMA ADPCM at 4 bits per sample — 16 kHz mono at 8 KB/s per talker, a quarter of raw PCM and less than one moving enemy costs the netcode. Each packet carries its own predictor state, so a dropped packet costs 40 ms and nothing after it. Forty lines from a 1992 spec instead of a third-party library, which keeps mic selection in-game and the credits screen clean.
• Enemy counts budgeted by netcode, not by fantasy. Measured at ~31 KB/s per remote client for 100 moving enemies, so the design target is 60, the cap is 100, and they never spawn in one frame.
• Networked and seeded from the first commit. All randomness sits behind one seeded, server-authoritative service — desync is the number one failure in a co-op roguelike — with content as data (ScriptableObjects and JSON, stable GUID IDs) and a versioned save schema carrying a migration hook from the first write.