/curriculum vitae
Berk Kırık.
Senior AI Engineer with 5+ years building production AI in microservices — LLMs, RAG, agentic AI, and the Kubernetes infrastructure underneath. Background spans fintech and healthcare.
Experience
Sep 2025 — present Türkiye
Senior AI Engineer
Production LLM and agentic-AI systems on a microservices architecture.
- Code-Generation Fine-tuning. Used LoRA on Hugging Face to fine-tune large code models, boosting domain-specific code-completion accuracy by 40% while cutting training cost by 90%. Only 0.1% of model parameters needed tuning, with A/B testing for seamless adapter swapping.
- Multi-Tenant NER System. Scalable Named Entity Recognition platform serving multiple enterprise clients with tenant-specific isolation, automated model selection based on SLAs, and sub-200ms inference latency.
- Guardrail System. Multi-layer guardrails (Guardrails AI + NeMo + custom) blocking 89.2% of harmful outputs. PII detection at 97% accuracy across 15+ categories, SHAP explainability, full-stack monitoring app (React / FastAPI / PostgreSQL), Redis caching for sub-50ms responses, Prometheus / Grafana observability — handling 5K+ concurrent requests.
- Enterprise RAG System. Production pipeline processing 500K+ documents with hybrid search (vector + BM25), advanced chunking, HyDE query rewriting, cross-encoder re-ranking, and RAGAS evaluation. 92% recall@10, 45% reduced hallucinations, sub-3s responses, 20+ file formats.
- Agentic AI Development. Multi-agent orchestration with LangGraph / CrewAI (ReAct + Plan-and-Execute) across 25+ tools. Autonomous code-review agents, multi-agent collaboration (researcher / analyst / writer / critic), episodic + knowledge-graph memory, LangSmith tracing, self-healing pipelines at 99.2% task completion, guardrail integration.
Sep 2023 — Sep 2025 Frankfurt am Main, Germany
Data Scientist Team Lead / Backend Engineer
Led the ML pipeline lifecycle for financial-message compliance, fraud detection, and a cloud-native banking platform.
- ML Model Development & NLP Automation. Led the full ML pipeline lifecycle for validating ISO 20022 and SWIFT MT/MX financial message formats. Built NLP-driven systems in Python to automate compliant message generation.
- LLM Fine-Tuning, Generative AI & RAG Integration. Directed LLM fine-tuning for compliance, fraud detection, and message classification (Hugging Face, PyTorch). Integrated Generative AI for automated message construction and domain chatbots, RAG pipelines combining FAISS vector stores with LLMs for regulatory support.
- Graph-Based Fraud Detection & Data Mining. Graph-based systems on Neo4j + GraphQL — link prediction, clustering, anomaly detection — combined with Pandas / Spark mining for real-time signals over large financial datasets.
- Backend Platform Engineering (Golang + Python). Designed and led a cloud-native banking platform: Golang for performance-critical services, Python / Django for AI orchestration. Globally distributed Agile teams.
- Microservices Architecture with gRPC & REST. Production microservices (Golang + Python). gRPC for internal, REST for external. KrakenD as API gateway, Keycloak for RBAC + ABAC.
- Kubernetes Infrastructure & DevOps Automation. Kubernetes-native infrastructure with custom Golang operators / controllers for dynamic orchestration and autoscaling. Shell scripts, cron jobs, GitOps-driven CI/CD pipelines.
- Agile Team Leadership. Lead for the Data Science team across multiple Agile squads — sprint planning, retrospectives, cross-team collaboration with engineering and product.
Aug 2021 — Sep 2023 Ankara, Türkiye
Data Scientist
Healthcare AI — sensor-fusion activity recognition from edge to mobile.
- Device & Sensor Integration. Multi-device sensing pipeline using ESP-32 WiFi modules and Raspberry Pi to gather real-time motion, environmental, and location-based sensor data — for continuous edge monitoring.
- Data Processing & Synchronization. Timestamp synchronization, noise filtering, and data normalization (Min-Max, Z-Score) for high-quality input to downstream ML. Handling for missing data, sensor drift, and latency.
- Feature Engineering & Dimensionality Reduction. PCA, FFT, and LDA for feature extraction — improving signal separation between user activity classes, model interpretability, and computational overhead.
- Machine Learning & Deep Learning Models. Decision Trees, SVM, Random Forest, and CNNs for activity classification from sensor streams. XGBoost and ensemble approaches for generalization and reducing overfitting.
- Backend API Development. FastAPI for real-time inference APIs, Django for user data, authentication, and application logic. RESTful endpoints supporting mobile and web apps, with logging, feedback loops, and results visualization.
- Mobile App Deployment & Real-Time Feedback. Mobile app interfaces embedding trained models for real-time alerts, personalized activity tracking, and adaptive feedback. Optimized model size and latency for on-device inference.
- Model Fine-Tuning & Performance Optimization. Transfer learning, hyperparameter tuning, and data augmentation to reduce false positives. Refined CNN and NLP-based architectures for healthcare constraints and variability.
- Validation & Evaluation Techniques. k-fold cross-validation, Grid Search, and Random Search to systematically evaluate and optimize model performance across diverse users and activity scenarios.
- Graph-Based Behavioral Analysis & Data Mining. Neo4j + GraphQL + Pandas to model and query user behavior patterns through graph analysis. Actionable insights from streaming sensor data for healthcare and activity monitoring.
Education
Feb 2026 — present Ankara, Türkiye
MSc, Computer Engineering
Ankara University
Aug 2018 — May 2024 Ankara, Türkiye
BSc, Biomedical Engineering
Aug 2022 — Jul 2024 Istanbul, Türkiye
Associate, Computer Programming
Skills
Programming languages Python · Go · C++ · PHP · SQL
ML / Deep Learning PyTorch · TensorFlow · Keras · JAX / Flax · Hugging Face · LoRA · XGBoost
LLM / agentic systems LangGraph · CrewAI · LangSmith · RAG (FAISS, BM25, HyDE) · RAGAS · Guardrails AI · NeMo Guardrails
Backend & APIs FastAPI · Django · Flask · Node.js · gRPC · REST · KrakenD
Data & graphs PostgreSQL · Oracle · Redis · Kafka · Neo4j · GraphQL · Pandas · Spark
Infra & DevOps Kubernetes · Docker · Docker Swarm · GitOps · AWS · Prometheus / Grafana · Keycloak
Biomedical / domain Signal processing (FFT, PCA, LDA) · Sensor fusion · Embedded systems (ESP-32, Raspberry Pi) · Healthcare AI · Biological data / classification
Publications
2026 Springer — The Journal of Supercomputing
Berk Kırık, Güney Uğurlu, Ayhan Aydın, Fatih Ekinci, Koray Açıcı, Eda Kumru, Aras Fahrettin Korkmaz, Mustafa Sevindik, Mehmet Serdar Güzel, Ilgaz Akata
First author. Hybrid CNN–KAN architecture for fine-grained classification of 8 Lactarius mushroom species (1,614-image curated dataset). 98.77% accuracy / 0.9994 AUC — outperforming ConvNeXt-Small, EfficientNetV2, ResNet-50, MobileNetV3, RegNetY, SqueezeNet, and ViT-Small. Friedman + Nemenyi tests confirmed significance. LIME-based XAI showed correct predictions relied on biologically meaningful structures (gill lamellation, cap zonation, stipe–cap transitions).
2025 Nature — Scientific Reports
Yasin Atilkan, Berk Kırık, Eren Tuna Açıkbaş, Fatih Ekinci, Koray Açıcı, Tunç Aşuroğlu, Recep Benzer, Mehmet Serdar Güzel, Semra Benzer
KANs at 95–100% accuracy on tabular features; stacked-autoencoder + KAN architecture lifted image-feature performance ~3.5%. McNemar / Wilcoxon tests confirmed significant differences vs. SVM, MLPs, and naive Bayes.
2024 Peer-reviewed journal
Advancing Crayfish Disease Detection: A Comparative Study of Deep Learning and Canonical Machine Learning Techniques
Yasin Atilkan, Berk Kırık, Koray Açıcı, Recep Benzer, Fatih Ekinci, Mehmet Serdar Güzel
Hybrid RF-ResNet50 (ResNet50 features + Random Forest classifier) outperformed both pure deep-learning and canonical-ML models on an imbalanced dataset.
Full record on ORCID.
Interests
Open-source contributions · Swimming · Basketball · Chess puzzles