Personal craft, architectural sandboxes, and software built purely out of curiosity.

github.com/Evgenii-Zinner/Sting

A 2D ECS-first game engine built in Dart. Because clearly, the software world was suffering from a tragic shortage of game engines.

The Rendering Thesis: Instead of building on top of Flutter or relying on Flutter-based engines like Flame, the experiment was to write directly against raw dart:ui. The goal: identical visual rendering across platforms without the overhead, layout passes, or constraints of Flutter’s widget tree.

Agentic Orchestration: Built as an end-to-end experiment in autonomous AI engineering. Jules was used for all development tasks—from curating subagent outputs and evaluating architecture to writing the exact engine code—orchestrated via the Antigravity CLI and executed directly against the Jules API.

Pure experimental sandbox. Tested the practical limits of raw Dart rendering pipelines and automated agent collaboration on non-trivial codebases.

github.com/Evgenii-Zinner/nodlets · nodlets.ezinner.com

An in-browser simulation of tiny digital creatures that intercept and hoard dropped network packets. Built with PixiJS v8 and deployed on Cloudflare Workers.

The Simulation Sandbox: A dense 2D canvas loop rendering thousands of autonomous entities navigating spatial fields. Rather than relying on abstract benchmark suites, the project creates an active visual playground with collision math, packet intercept mechanics, and tight frame-budget constraints.

The Agentic Terrarium: Built specifically as a live testing ground to observe how Google Jules and autonomous performance personas (like Bolt) behave in practice. I needed a disposable project with heavy mathematical throughput and genuine render-loop bottlenecks to watch how agents profile hotspots, inline loops, and chase micro-optimizations in the wild without risking production code.

Pure experimental playground. An interactive digital terrarium that doubles as a disposable stress-testing arena for watching autonomous coding agents work.

Closed Alpha · Voice-First Telegram Bot

A voice-first Telegram speaking partner designed to build real-world English fluency and speaking confidence. Built with Bun, grammY, Google Gemini, and Google Cloud Chirp 3 HD.

The Conversational Ergonomics Thesis: Most language learning apps trap users in gamified multiple-choice quizzes or halt speaking flow with pedantic, red-pen grammar corrections. Engy operates strictly voice-to-voice: you send Telegram voice notes, it transcribes and interprets messy speech, deduces accent distortions through conversational context, and corrects mistakes organically through natural recasting rather than interrupting momentum.

Adaptive CEFR & Cognitive Load Guardrails: Tailored specifically to eliminate beginner anxiety. For A1/A2 learners, audio playback rates are dynamically slowed down (0.55x–0.65x) using high-fidelity Chirp 3 HD voice synthesis, and replies enforce a strict two-sentence structure: one sentence to warmly confirm understanding with recast corrections, and exactly one focused follow-up question. Backed by a self-correcting SQLite memory that updates user background notes dynamically and automated scheduled voice prompts that check in on idle learners.

Pure conversational sandbox. Currently running in closed alpha to test acoustic naturalness, adaptive CEFR progression, and long-term learner retention.

Not all experiments make it to production. For killed projects, falsified hypotheses, and the reusable architecture salvaged from the wreckage, explore Scrap & Post-Mortems.