Machine learning & applied AI
Training and evaluation pipelines, plus LLM-backed product features — retrieval-grounded agents, voice-matched generation, and the guardrails that keep them safe to ship.
00 / Profile
Vienna, Austria
Independent
Machine learning / product engineering / interface design / web
I build systems and the products around them.
Independent engineer working across machine learning, full-stack product software, and the web. End-to-end delivery: data and model, backend and interface, design and shipped surface. One person, full ownership of the stack.
Education
BSc Artificial Intelligence
Johannes Kepler University Linz
Engineering range
Data · Model · Backend · Frontend
Python, PyTorch, TypeScript, React, Node, Postgres, Swift
Languages
English · German · Russian · Kazakh
Comfortable working with international teams
01
Capabilities
Four practices I bring to a project, solo or alongside a team.
I work end to end when a project needs that, and I plug into an existing team when that's the better fit. Comfortable picking up tickets in a sprint, pairing on a hard bug, or owning a feature from research to release.
Product engineering
Full-stack work on web apps, desktop tools, bots, and customer-facing surfaces. Equally happy implementing a spec or shaping one with the team.
Web & storefronts
Landing pages, marketing sites, and storefronts that hold up to brand scrutiny. Performance-minded, editable by non-engineers, friendly to handoff.
Interface & design systems
Interaction models, responsive layout, and design decisions that make technical products easier to use. Designs documented for the next person, not just shipped.
02
Selected work
Eleven projects across applied AI, product software, and brand-grade web.
Each case study is framed by the decisions that mattered: data and system structure, interface clarity, the safety rails around an AI feature, and the result the work had to deliver.
01
Applied AI / Desktop
SilkOffRoad AI Agent
Local-first macOS cockpit that drafts grounded, guardrailed replies to tour inquiries — retrieval over a master dataset, a dual-LLM pipeline, and Gmail sync.
2026
02
AI agent / Automation
LinkedIn Content Agent
Researches AI topics, drafts posts in my voice, gates each one through a Telegram review step, then auto-posts to LinkedIn — a human always in the loop.
2026
03
AI translation / Language
Soile
Offline-ready Kazakh–Russian–English speech translator and phrasebook with confidence scoring and a built-in language-learning mode.
2026
04
Product / Tour operations
SilkCRM
FastAPI + SQLite CRM for a tour business — tours, bookings, payments — with passport OCR that auto-fills the generated agreements.
2026
05
Web app / Print product
Nice Maps
A map studio that turns route planning into export-ready PNG/PDF artifacts — Mapbox GL rendering, controlled palettes, and print presets.
2026
06
E-commerce / Brand system
Origo Jewelry
Editorial Next.js storefront for an independent jewelry brand — a canvas ring animation, restrained merchandising, and a considered checkout.
2026
07
Landing page / Service business
Marina Psy
Conversion-focused landing for a psychologist — credibility-first layout, a blog, a socionics test, and Postgres-backed lead capture.
2026
08
Telegram bot / Fintech
MoneyTracker
A trip-expense Telegram bot with a local parser plus Gemini fallback, six-currency live conversion, per-tour Excel ledgers, and a running wallet.
2026
09
Native macOS app
Clockio
A SwiftUI/AppKit desktop clock, screensaver, and to-do app with multiple clock styles and an offscreen snapshot QA pipeline.
2026
10
Audio ML / Sound-event detection
Sound-event detection
Multi-label classification of 15 everyday sound events per second of audio — leakage-safe evaluation, then an AudioSet transfer-learning system reaching 0.70 macro-F1.
2026
11
Computer vision / Remote sensing
Satellite image classification
A from-scratch PyTorch CNN that sorts EuroSAT tiles into ten land-use classes at 96% accuracy — served through an interactive Shiny web app.
2026
03
Method
Build the system before polishing the surface.
I start by defining the workflow, the data shape, and the user decision that the interface must support. Visual design comes from that structure instead of being applied as decoration at the end.
InputData / brief / constraints
CoreModel / service / product logic
OutputInterface / deployment / iteration
04 / Contact
Currently accepting new engagements.
Best fit: product software, applied machine learning, brand-grade landing pages, internal tools, storefronts, and customer-facing systems where engineering depth and design are both part of the deliverable.