Tarraco Depths
Procedural PC game prototypeAtmospheric exploration prototype with procedural layouts, spatial audio, and a custom lighting pass.
Still non-playable
Unreal Engine 5 · PC
2025 · Prototype
On request
Software engineer for web platforms and applied AI.
I’m Zendryz. I design and ship production systems - TypeScript frontends, Node APIs, and LLM infrastructure - with an emphasis on measured performance and code others can maintain.
- 04+
- Years building and shipping
- 12
- Systems designed end to end
- 02
- Research tracks in progress
AreasWeb platformsLLM systems & agentsInference optimization
01 | Selected work
Work with measurable outcomes.
Open-source systems and research. Every entry links to code or to documented results - no mockups presented as shipped work.
No.ProjectStackYearLink
Stack Composer
CLI for production scaffoldsCommand-line tool that generates typed, linted, CI-ready project scaffolds instead of starter-template copy-paste.
Published on npm
Node.js · TypeScript
2025 · Shipped
KraftOS
Light-weight OSOperating system with custom-made kernel capable of running Windows applications
Everything works fine
C · Assembly
2026 · Private
On request
GSSO
Speculative inference researchSparse orchestration setup for LLM inference: pruned draft model plus asymmetric distillation to cut training cost.
−70% train cost · +45% throughput (local bench)
PyTorch · CUDA · Python
2026 · Research
Full archive -github.com/ZendrYz
02 - Profile
Profile.
Independent engineer, previously across web product and ML tooling work. I take on a small number of projects at a time.
I build the full system: interface, API, data layer, and - where AI is involved - the evaluation harness around the model. Most of my work sits in TypeScript and Python, and most of it is still running in production.
Current research time goes to GSSO, a speculative-decoding setup aimed at lowering inference cost.
P.01
Production over demos
Everything ships with tests, observability, and a runbook. If it only works on my machine, it doesn’t count.
P.02
Measured performance
Latency, cost per 1k requests, and bundle size are tracked from week one - not asserted in a slide deck.
P.03
Maintainable by others
Typed interfaces, documented decisions, boring technology where it fits. Handover is part of the deliverable.
Tooling I reach for
- Frontend
- React, Next.js, Astro, Tailwind
- Backend
- Node.js, PostgreSQL, Redis, Docker
- AI
- Python, PyTorch, LangGraph, RAG pipelines
- Ops
- CI/CD, error tracking, cost dashboards
03 — Services
Three engagements, clearly scoped.
Fixed scope, fixed price, weekly demos. If a project doesn’t fit one of these, I’ll say so on the first call.
S.01
Web platforms
Marketing site to logged-in product: typed frontend, API, database, auth, and billing wired correctly from day one.
Includes
- React / Next.js / Astro builds
- Node APIs + PostgreSQL
- Auth, billing, roles
- CI, previews, error tracking
01 / 03
How a project runs
01
Scope
Fixed written quote. What’s in, what’s out, what it costs.
02
Prototype
Clickable build in week one. Killed early if wrong.
03
Build
Weekly demos against the scope. No big-bang reveal.
04
Handover
Docs, runbooks, and a walkthrough. You own it after.
04 — Log
Devlog.
Build notes and measurements. Short entries, written when there’s something concrete to report.
gssoml
GSSO - Efficient Ghost Model Training
Managed to train the speculative ghost model at a fraction of the usual cost using dynamic layer pruning and asymmetric distillation. Promising throughput results.
I'd been going back and forth for weeks on how to cut down the ghost model training cost in GSSO without sacrificing speculation accuracy. The vanilla approach required training a full transformer (~7B) just to do speculative decoding - massive resource waste for a model that only verifies predictions. The fix: dynamic layer pruning + asymmetric distillation. Instead of training a full model, I train the ghost model with a dynamically pruned architecture: early layers are lightweight (low-rank attention), only the last 4-6 layers keep full resolution. Asymmetric distillation forces the ghost model to learn speculative acceptance patterns only, not the full distribution. Results: - Training cost down ~70% - Inference throughput +45% vs. standard speculative decoding - Acceptance rate dropped only 0.03 (0.92 → 0.89) - negligible given the savings Next up: implement dynamic tree drafting with this lightweight ghost model.