Ismail Adiyaman

What I built

• Owned the game’s look end to end under a 72-hour deadline: lighting, animation and its implementation, level design, and asset integration.
• Lit the levels so the three cube scales stay readable at a glance. The mechanic only works if you can instantly tell which size you are.

How

• Godot. Every visual decision was made against the clock, so the look had to come from lighting and composition rather than asset volume.

Role

What I built

• A material that takes two custom textures and splits them across a surface (top half from the first, bottom half from the second), so the whole game could be themed fire/water fast and coherently.
• Tree and bush shaders tuned to sit inside that art direction rather than fight it.
• The flame VFX, and a contribution to the water VFX.
• 3D assets and skybox optimization.

How

• Unity, authored as custom shaders instead of per-asset texture variants, so one material change re-themes every surface in the game.
• Effects built against the co-op camera framing, so both players read their own element at a glance.

Role

What I built

• Built the background environment art and its animations for the build that shipped to gamescom.
• Produced that art under strict performance budgets so it held frame rate on the exhibition floor.

How

• Unity. Constraint-aware art: every layer costed against the frame budget of the machine it would actually run on.

What I built

• Lead the project across design, visuals, and programming: scope, schedule, and the direction the team works to.
• Main programmer on the loop that ties it together: gathering, crafting, the potions that change how the outskirts play, and the economy that sells them.
• Author the shaders and tech art that give the isometric world its handmade look.
• Set and hold the visual direction so the work of several people reads as one world.

How

• Godot, with the rendering work written directly in its shading language rather than assembled from stock nodes.
• Decisions documented so a small, part-time team can move without me in the room.

Role

What I built

• Built the background environment art and its animations for the build that shipped to gamescom.
• Produced that art under strict performance budgets so it held frame rate on the exhibition floor.

How

• Unity. Constraint-aware art: every layer costed against the frame budget of the machine it would actually run on.

What I built

• Lead the project across design, visuals, and programming: scope, schedule, and the direction the team works to.
• Main programmer on the loop that ties it together: gathering, crafting, the potions that change how the outskirts play, and the economy that sells them.
• Author the shaders and tech art that give the isometric world its handmade look.
• Set and hold the visual direction so the work of several people reads as one world.

How

• Godot, with the rendering work written directly in its shading language rather than assembled from stock nodes.
• Decisions documented so a small, part-time team can move without me in the room.

Role

What I built

• The block model: 144 ten-minute blocks a day, with nestable activities three levels deep, painted by tap or drag and erased by tapping again.
• A clock-driven overlay that renders past blocks solid, animates the current block as it fills over its ten minutes, and leaves the future faint, while keeping every block editable.
• Server-side SQLite through Drizzle with generated migrations, so the same plan is live on the phone and the laptop.
• Installable as a PWA, with unit tests over the block logic.

How

• The whole design serves one rule: assigning an activity has to be as close to zero effort as possible. The app is a tracker, not somewhere to spend time.
• SvelteKit with adapter-node, Drizzle for schema and migrations, Vitest for the logic that would silently corrupt a day if it broke.

Role

What I built

• The full voice path: microphone → Silero VAD endpointing → parakeet-mlx speech-to-text → a deterministic NLU router → skills.
• A daemon that warms the STT model on start and exposes a streaming SSE endpoint, so the HUD shows the transcript live as it is heard.
• A permission gate for any skill that touches the network: the first call asks, and the grant is scoped to the session.
• A dependency-free CLI path on the standard library alone, plus a test suite over the routing and skill layers.

How

• Deterministic routing instead of an LLM: the same phrase always does the same thing, and it works with the network off.
• Four rewrites (v1 → v4). v1–v2 leaned on models and were unpredictable; v4 is the version that earned its place in my daily workflow.

Role

What I built

• A retrieval-augmented memory layer so the assistant kept context across sessions instead of starting cold every time.
• macOS integration: launching and controlling applications, media and volume control, Spotify, and Alfred-triggered search.
• Web search with content extraction feeding back into the model’s context.
• A test suite covering the interactive, chat, and integration paths.

How

• Everything local and privacy-first: models run through Ollama on the machine, nothing is sent out unless a search is asked for.
• Building it is what convinced me a model-driven assistant is the wrong shape for a daily tool, because the same request would not reliably do the same thing. Jarvis is the deterministic answer to that.