Retrieval-augmented applications
Python applications that connect retrieval, metadata filters and model calls, with checks on the evidence used to answer a question.
Katerina Zhittsova · AI engineer
I build AI applications and put their answers to the test.
My focus is retrieval-augmented generation (RAG), guardrails and LLM evaluation. I want useful answers with evidence behind them, and clear behavior when that evidence is missing.
RAG & LLM Evaluation · Python & SQL · Data Engineering
Featured AI project
A learning assistant with layered guardrails
A learning assistant should help you work through course material and show which sources support its answer. I built this project to compare baseline RAG with layered guardrails: where each check helps, where it rejects a valid request and whether the answer has enough evidence behind it. The pipeline uses BGE-M3 retrieval, Chroma filtering and OpenAI-compatible classification, generation and verification.
You can try an offline demo, follow a request through the pipeline, or inspect the tools for reviewing evaluation data. The evaluation workflow keeps calibration separate from final evaluation.
What I work on
LLM evaluation and guardrails
Baseline comparisons, answer verification and abstention behavior. I build review tools and keep calibration decisions separate from final evaluation.
Data engineering
Python and SQL pipelines, ingestion debugging, validation checks and reporting datasets, including spatial analysis with PostgreSQL/PostGIS.
The data background behind the AI work
My professional background is in analytics and data-engineering workflows for energy, renewables and agribusiness. I've worked with operational, environmental and GPS data, from investigating broken ingestion to building datasets for reporting. I bring that attention to sources and validation into my AI applications.
I'm based in Berlin and looking for AI Engineer roles. I'm also interested in data engineering within AI teams. More about me.
Highlights
Latest blog
Evaluating ML, deep learning, LLMs and reinforcement learning
20 Sept 2026
How to choose the evaluation unit, data split and metrics for predictions, generated answers and sequences of actions, with worked examples and a RAG case study.
Featured case study
Agri-Weather-Yield Drivers
6 Apr 2026
A weather, yield, soil, and geospatial analysis project for risk-aware siting and environmental review.
Pinned projects
- Crude Oil Benchmark Analytics
Built the first public crude benchmark visual bundle, including interactive API/sulfur and regional context views.
- agri-weather-yield-drivers
Built a rerunnable analytics pipeline for risk-aware siting conversations, with transparent assumptions and map-ready layers.
Browse the full project list on /portfolio.
How I work
I like following a result back to its inputs, whether it's an LLM answer or a reporting dataset. That means making the system easy to inspect as well as useful to run.
- Start with the question someone needs answered and define what a useful result means.
- Build a baseline before adding complexity, then compare behavior on the same cases.
- Inspect source coverage, transformations and failures, including unsupported answers.
- Keep tests, evaluation notes and run instructions close to the code.