Stanko Kuveljic

Designed the AI and relevance layer for a real-time news application. The system combines Qdrant-based vector retrieval, personalization signals, and custom ranking to construct user feeds. I also developed a news clustering pipeline using semantic embeddings, time decay, and keyword extraction, together with evaluation workflows for LLM-based summarization and clustering.

At Vantage Discovery, later acquired by Shopify, I built an enterprise-grade semantic search and reranking platform for large-scale eCommerce product catalogs. I implemented modern embedding models, custom rerankers, and automated experimentation workflows using Databricks and MLflow to accelerate search testing and optimize inference scalability.

I designed a graph-based RAG and multi-agent reasoning architecture to automate complex US CFR regulatory compliance workflows. I orchestrated regulation extraction, applicability reasoning, and product-specific checklist generation using Neo4j, LangGraph, and Anthropic Claude on Amazon Bedrock. The system achieved 91% recall in identifying applicable regulatory sections and reduced manual compliance review times by over 70%.

During my tenure at OLX, I developed and optimized advanced machine learning infrastructure, focusing specifically on vector embedding and hybrid search capabilities. I automated a fine-tuning pipeline utilizing MLflow and AWS SageMaker. This robust infrastructure enabled highly parallelized training, allowing the team to scale experiments and drastically reduce overall experimentation time.

To further improve efficiency, I executed a deep optimization of model training on T4 GPUs. By applying advanced memory management techniques to minimize VRAM and CPU usage, I successfully condensed model training time from 20 hours down to just 20 minutes. This massive performance leap simultaneously drove a 90% reduction in associated AWS compute costs.

Finally, I deployed these assets to production, setting up the hybrid search models as highly scalable APIs. This was achieved by leveraging a centralized model registry alongside LitServe, orchestrated within a Kubernetes environment to ensure reliability and speed.

I collaborated with the system architect to design and build an AI-native conversational platform for a music school. The primary focus was on engineering a multi-agent scheduling bot to replace static forms and drastically speed up administrative workflows like booking lessons, managing professor availability, and rescheduling time slots.

I designed and implemented a deterministic, state-driven workflow using LangGraph, combining slot-filling mechanisms with dynamic UI elements and free-text inputs. I built a multi-agent system using OpenAI models and custom tools capable of handling both domain-specific Q&A and transactional scheduling operations. I also integrated FastAPI and PostgreSQL for the back-end service layer, alongside Langfuse for LLM observability and performance evaluation.

I served as tech lead for a team of 3, engineering an enterprise Next Best Action (NBA) machine learning platform for pharmaceutical sales representatives during targeted medication campaigns. The system predicts optimal omnichannel engagement strategies—such as direct calls, emails, or SMS—to maximize outreach efficiency with healthcare professionals.

I designed the core NBA architecture, built feature store pipelines in Snowflake, and established automated MLOps pipelines using AWS SageMaker for continuous model training and daily scheduled batch inference. To tackle low-engagement signal sparsity, I engineered recency-weighted interaction aggregations and cross-channel composite features. Implemented XGBoost models that provided daily actionable outreach recommendations directly to sales teams.

As a forward deployed engineer at NextData, I worked on proof-of-concept (POC) engagements for enterprise clients to demonstrate the value and feasibility of decentralized data mesh architectures. I worked directly with client engineering and domain teams to translate legacy, centralized data pipelines and complex business use cases into autonomous, domain-owned data products on the NextData platform.

I designed and implemented end-to-end data product prototypes leveraging Databricks for scalable data processing and Apache Kafka for real-time data streaming. I also developed custom integration scripts and data pipelines in Python to connect client infrastructure to the platform, enabling self-serve data sharing, MCPs, and automated federated governance across business units.

I worked as a back-end developer on a workplace analytics platform that processes spatial data to deliver real-time space utilization insights for corporate real estate. The system unifies raw telemetry from diverse sources—including Wi-Fi network signals, physical IoT occupancy sensors, and room calendar booking systems.

I contributed to scalable back-end services and high-performance API endpoints using Python and FastAPI. I improved spatial count accuracy by implementing data cleaning and processing algorithms with NumPy to filter signal noise and harmonize discrepancies between raw sensor counts and active bookings. I also built integration pipelines to ingest external sensor data feeds and managed scalable data storage pipelines using AWS S3.

Engineered a computer vision space analytics platform processing video feeds from overhead ceiling-mounted fish-eye camera sensors. Fine-tuned custom YOLO-based object detection models specifically optimized for fish-eye lens distortion, achieving 95%+ detection accuracy for people and physical assets. Developed a multi-object tracking algorithm to reliably assign and maintain individual IDs across camera frames.

Built dual real-time and batch processing architectures. Implemented serverless inference on AWS Lambda for live stream processing, real-time occupancy counting, and automated event alerting. For historical analysis, built data processing pipelines using AWS Batch, S3, DynamoDB, and Amazon Athena to power custom Power BI reports of space analytics. Standardized MLOps workflows using MLflow and LakeFS for data and model versioning, accelerating onboarding time for new physical sites.