ZOOP · Creator network
Personal feeds for 10,000+ monthly users
10K+monthly users get a personal feed
Recommendations · ML · AWS
I build AI systems
Tell me about your projectthat run in production.
I'm Artem Zelenskii, an AI/ML engineer in Vienna. For the last 5 years I have built recommendation engines, RAG pipelines and AI agents, from the first data pipeline to live inference.
- Recommendation systems
- RAG and LLM pipelines
- AI agents and automation
- Data pipelines and ML infrastructure
Selected Work
Cases with numbers
About
Who you work with.
I'm Artem, an AI/ML engineer based in Vienna. I work in German, English and Russian.
For the last 5 years I have built the AI and machine learning side of applications. At ZOOP I was the data science developer and recommendation system architect. I built the recommendation engine from scratch: the signal pipeline, retrieval, ranking, cold-start rules for new users and creators, and scheduled retraining on AWS. Today 10K+ monthly active users open a feed ranked for them.
Since then I have built recommendations for a top-15 global grocer, an AI lead-scoring CRM for a B2B SaaS company, a chess tournament platform listed by the German Chess Federation, and AI agents that run content and social media routines on their own.
How I work
Every agent I build starts by assisting a person, then drafts work for a named person to approve. It acts on its own only for narrow, low-risk tasks you sign off on, with a full log of what it did.
- 01
Map the work
I sit with the people who do the work and map how it really runs today, workarounds included.
- 02
Design the guardrails
With you, I pick the processes that are ready to automate and set approval steps for each one before I build anything.
- 03
Build and prove
I build against real cases from your business and test against a held-out dataset with pass bars set in advance.
- 04
Connect and run
The system joins the tools your team already uses, with cost per run, spend caps and an alert to a person when a run fails.
- 05
Grow and hand over
I extend what the system does on its own only as the evidence supports it. You own the full code and data of your version.
What I do
Specialties
Recommendation systems
Feeds and catalogs that feel personal from the first session. Retrieval, ranking, cold start and retraining.
- Python
- PyTorch
- FAISS
- OpenSearch
RAG and LLM pipelines
LLMs that answer from your own data or score your leads, each decision with a written reason, tested on held-out data before launch.
- Gemini
- OpenAI API
- LangChain
- PostgreSQL
AI agents and automation
Agents in n8n and Telegram that handle daily routines, from video-to-article pipelines to running 100+ Instagram accounts from one chat.
- n8n
- Telegram
- Gemini
- Apify
Data pipelines and ML infrastructure
Event pipelines, Airflow ETL and model serving on AWS, built to re-run safely from stored raw data.
- Airflow
- AWS
- Polars
- FastAPI
What I build with
Tech stack
The tools I reach for by default. The right stack still depends on what you are building, so I treat this as a starting point, not a template.
Machine learning
Recommendation and ranking models, from retrieval and cold-start rules to scheduled retraining.
- Python
- PyTorch
- TensorFlow
- scikit-learn
- XGBoost
- FAISS
RAG and search
Embeddings, dense and keyword search and re-ranking, as one search layer the whole system shares.
- OpenSearch
- Hugging Face
- PostgreSQL
LLMs and agents
Agents with one narrow role each and their own tools. I pick the model on merit and keep it switchable.
- Claude
- OpenAI
- Gemini
- LangChain
- n8n
Data and MLOps
Scheduled pipelines and retraining on AWS, packaged the same way from laptop to production.
- AWS
- Airflow
- Docker
- Polars
Apps and APIs
Python APIs next to the ML layer, so logic lives in one language, and dashboards your team works in.
- FastAPI
- React
- Next.js
- TypeScript
- Supabase
Whatever the tools
Queues
Every handoff between processes goes through a durable queue. If a consumer is down, the task waits instead of vanishing.
Guardrails
I scope every agent action in advance. Sending email, moving money or deleting a record sits behind an approval step by default.
Data security
Secrets live in a vault, never in the repo. Access is scoped per service, data is encrypted in transit and at rest, and sensitive actions are logged.
Let's talk
Have a system that needs to work in production?
Tell me what you want to automate or personalize, and what data you already have. I'll reply with how I would approach it.
Vienna, AustriaGerman, English, Russian