Multimodal search & analysis platform
Image and text retrieval over millions of enterprise assets. Took it from a PoC that couldn't ship to a stable multi-client product, deployed both on-prem and in the cloud.
FastAPI
Celery
RabbitMQ
Docker
01
What I actually built
Vector search that costs 90% less
Moved retrieval from an in-memory HNSW index (Weaviate) to PostgreSQL with pgvector and DiskANN, and lowered the embedding dimensionality without significant quality loss. Result: the AWS bill fell by 90%.
PostgreSQL- pgvector
- DiskANN
- Weaviate
Solr
Multi-stage computer-vision pipelines
Object detection and region-of-interest extraction for print and packaging artwork, tuned for high recall. Includes a rule-based surface inspector that replaced an AI model: over 10× faster (7.5 s → 0.6 s per sheet), no GPU needed to run on-prem.
PyTorch- OpenCLIP
- DINO
OpenCV- OCR
Pixel-exact search across millions of images
Finds designs that differ from the query by just a few pixels, down to a single one: the same packaging without a tiny technical mark, or with one small extra logo. The search takes seconds; comparing every pixel against every file in the archive directly would take hours to days.
- OpenCV
- PostgreSQL
- pgvector
Secure, observable, in the cloud
JWT authentication, three access roles and rate limiting; Prometheus and Grafana monitoring; AWS EC2 deployments with CloudFormation; a bash installer/updater that sets up or upgrades a full customer deployment, and a CI/CD pipeline that builds all Docker images per release, plus an automated test pipeline covering every API endpoint.
JWT
Prometheus
Grafana
AWS
GitHub Actions
Close to the customer
Requirements analysis, integration workshops and hands-on rollout with clients and internal stakeholders, because a system only counts once someone uses it.
Full case study →02
Experience
11/2023 – todayVreden, Germany
AI Engineer (Computer Vision)
- Designed, developed and deployed a production-grade AI-based multimodal search and analytics platform (image and text retrieval), turning a non-production PoC into a stable enterprise solution: >99.9% uptime, multi-tenant.
- Built an asynchronous backend that processes large-scale production image and text data (millions of assets) while keeping the search, indexing and analytics APIs responsive.
- Developed multi-stage computer vision pipelines for enterprise image data with high recall requirements and dense object distributions.
- Implemented a unified hybrid search that combines metadata-based, full-text, semantic and visual retrieval in a single system.
- Moved retrieval from an in-memory HNSW index (Weaviate) to PostgreSQL with pgvector and DiskANN, and lowered the embedding dimensionality without significant quality loss. Result: the AWS bill fell by 90%.
- Enabled cloud deployments by working closely with an integration partner and implementing configurable AI worker profiles that balance compute usage and adapt to changing customer priorities.
- Integrated OCR-based text extraction for industrial image data to enrich enterprise search workflows with visually embedded metadata.
- Implemented production-grade security (JWT authentication, role-based access control) and observability to ensure system stability and availability.
- Built a CI/CD pipeline that builds all Docker images per release, plus an automated test pipeline covering every API endpoint.
- Worked closely with customers and internal stakeholders through workshops, requirements analysis and production deployments in on-prem and cloud environments.
Python · FastAPI · Celery · RabbitMQ · Uvicorn · PyTorch · Hugging Face · OpenCV · LLMs · OCR · PostgreSQL (pgvector, DiskANN) · Weaviate · Solr · Docker · AWS · GitHub Actions · Prometheus · Grafana · JWT
03/2022 – 11/2023Germany
Relocation & self-study
- Moved to Germany and completed a German course up to B1.
11/2021 – 02/2022Kharkiv, Ukraine
AI Engineer
- Built and trained end-to-end computer-vision pipelines, including image-captioning and image-segmentation models.
- Implemented encoder-decoder architectures with pretrained CNN backbones and attention-based sequence models.
- Set up preprocessing, training, evaluation and inference pipelines with reproducible experiments and performance logging.
- Evaluated models with NLP and CV metrics (BLEU, ROUGE) and visualised results in TensorBoard.
Python · PyTorch · torchvision · PIL · TensorBoard
03
Education, languages & toolkit
Education
B.Sc. Computer Engineering
National Aerospace University, Kharkiv, Ukraine
09/2013 – 06/2017
Languages
- German C1
- English C1
- Ukrainian Native
Toolkit
- Python
- PyTorch
- Computer Vision
- Machine Learning
- OpenCV
- Hugging Face
- LLMs
- OCR
- FastAPI
- Celery
- Uvicorn
- RabbitMQ
- PostgreSQL
- pgvector
- Weaviate
- Solr
- Docker
- AWS (EC2, CloudFormation)
- GitHub Actions
- Prometheus & Grafana