Matej Kalc · Portfolio
Full-Stack AI Engineer building LLM products from prototype to production.
I’m based in Trieste, engineering AI at NLB (opens in new tab) and founding Timeo (opens in new tab). I balance model quality, latency, and cost under real business constraints.
- RAG and evaluation
- Agent orchestration
- LangGraph workflows
- LLMOps for GenAI
Timeline
2025 - present
Full-Stack AI Engineer
NLB d.d., Ljubljana (opens in new tab)
- Led an agentic document-validation system with orchestration and checkpoints, reducing manual work by 1,000+ hours per year.
- Shipped an AI report generator with web search, retrieval-augmented generation, and image generation.
- Built dynamic graph neural networks for anti-money-laundering monitoring in banking workflows.
2025 - present
Founder
Timeo (timeoschedule.com) (opens in new tab)
- Built and launched Timeo, a scheduling SaaS for small businesses.
- For pilot users, schedule preparation dropped from about 3 hours to around 20 minutes.
2024 - 2025
Data Scientist
Zurich Insurance Group, Ljubljana (opens in new tab)
- Introduced graph neural networks for automobile insurance fraud detection and improved production performance by 20% in one business unit.
- Developed Databricks workflows that reduced batch inference time by 75%.
2023 - 2024
Data Scientist (Student)
Medius.si d.o.o., Ljubljana (opens in new tab)
- Created a Python package for solar power production forecasting.
- Improved one-day-ahead forecast quality by 25% with recurrent neural networks.
Selected work
Each project highlights context, technical approach, and measurable outcomes.
Valori Immobiliari Trieste
2026Interactive map of Trieste real-estate values with OMI zone ranges and local market filters.
Outcome
Turned public valuation data into a clearer browsing experience for exploring price differences across the city.
- Next.js
- Interactive Maps
- Real Estate Data
Timeo (SaaS) - Scheduling in seconds
2025Optimization-based scheduling platform that builds compliant schedules in seconds.
Outcome
Pilot users reduced schedule planning from about 3 hours to around 20 minutes.
- Optimization
- Product
- Operations AI
Automobile Insurance Fraud Detection
2024Heterogeneous graph neural network pipeline over claims, customers, and provider relationships.
Outcome
Improved production model performance by 20% in one business unit.
- Graph Neural Networks
- PyTorch Geometric
- Insurance ML
Traffic Prediction with Temporal GNNs
2023Sensor-graph forecasting model for highway traffic prediction several hours ahead.
Outcome
Delivered accurate spatiotemporal forecasts from real transportation sensor data.
- Temporal GNN
- Time Series
- Spatial Data
CTR Prediction at Scale
2022Compared neural and factorization models for click-through-rate prediction on Outbrain data.
Outcome
Built and tuned multiple model families with Bayesian optimization on HPC resources.
- CTR Modeling
- Deep Learning
- Bayesian Tuning
About
I build end-to-end AI products for production use, from orchestration and evaluation on the backend to usable frontend experiences. In regulated environments, I prioritize observability and reliability. In products, I focus on user value and iteration speed.
- Core stack
- TypeScript · Next.js · Python · FastAPI · PostgreSQL
- Modeling
- RAG · Agent Orchestration · LangGraph Workflows · LLM Evaluation · Tool-Using AI Systems
Achievements
- 1st place - IEDC Case Study Competition (NLB team, advancing to international round) (opens in new tab)
- 1st place - Data Science Hackathon (Etrel, EV charging clustering)
- 1st place - NLP word sense disambiguation with SloBERTa (opens in new tab)
- 3rd place - Data Science Project Competition (2021 and 2022) (opens in new tab)
Let’s talk
Happy to discuss AI engineering roles, product collaborations, or honest feedback on an LLM system you’re building.
Trieste, Italy