About
I build AI that reads data the way language models read text. My work is Data Language Models, tabular foundation models, and real-time AI systems, taken end to end from research and architecture to production.
Continual learningAgentic AIApache KafkaApache FlinkPublications
Tabular foundation model: what I built
Schema-2 is the tabular foundation model I built end to end: architecture, pre-training data and objectives, evaluation, serving, and the platform around it. It has more than 1.5 billion parameters and uses axial attention over rows and columns, a Perceiver-style global path, and a schema path. It reads a raw table directly, with no serialization, no preprocessing, and no column names, and one checkpoint of about 2.18 billion parameters carries every capability below. Read the research, the solutions, the architecture, or the overview.
Row-level prediction
Achieved. First on the classification and binary boards at every board size I tested, with one default configuration against rivals that were tuned for hours per dataset. The overall board became valid only after I found and fixed a harness error. Regression and multiclass stayed behind the leaders, and I say so.
Continual learning
Achieved. The only model in the comparison that takes on new tasks without a full retrain. The recipe that stayed is experience replay with 48 raw rows per task, after I found and fixed an adapter that was being re-created for every task.
Anomaly detection
Achieved. First among thirteen methods by both average rank and AUROC. Three model-based channels are combined by a label-free pseudo-anomaly selector, so no labels are needed to choose how to fuse the signals.
Probabilistic regression
Achieved. The lowest median CRPS and the best mean rank on the merged board of 102 regression datasets and 50 models, from one sealed configuration. A 999-quantile head with split-conformal calibration fixed the coverage weakness, and I re-ran a weakened rival fairly.
Metadata independence
Achieved. The only model in the comparison whose accuracy does not fall when column names are removed or replaced with noise. Models that read headers lose roughly seven AUC points when headers vanish.
Blind sector identification
Achieved. A new capability: name the industry sector of an unseen dataset from raw cell values alone, with column names zeroed and no labels. No earlier tabular model performs this task. Pair correlation between columns was the signal that separated neighbouring sectors.
Missing-value imputation
Achieved. First place on the real-data board, ahead of language models and the classical imputers in every setting, with the widest margin under not-at-random missingness. Categorical imputation has no published competitor.
Entity and schema matching
Achieved. Beats the published state of the art on the entity matching benchmarks with no keys, and beats every name-aware matcher on all four schema scenarios while running name-blind.
Missing-data robustness
Achieved. A mean AUC above the published method across the full grid of 750 settings. It is a narrow lead and I report it as one.
One checkpoint, then compression
Achieved. Every capability merged into one unified checkpoint of about 2.18 billion parameters, with the code embedded and one line of bootstrap. The model was dumped at half, 8-bit, and 4-bit precision and checked for fidelity, and the model card was published after a contamination audit.
Timeline
The ideas and systems I have worked on, newest first. Orange marks publications, blue marks areas of work. Hover over any tool to see what it is used for, and select it to read more.
A new model class for tables: the Data Language Model
A foundation model that reads raw tables directly, with no serialization and no preprocessing. The paper defines the category. Read the paper, the research, the solutions, and the benchmark timeline that shows how each capability was built.
Foundation model for structured data
- Axial attention over rows and columns, Perceiver IO cross-attention, masked cell modeling for self-supervised pre-training, neural imputation, semantic classification heads, and memory-augmented continual learning.
- Large-scale synthetic data, distributed GPU training, mixed precision, evaluation on public benchmarks, and production serving.
- Training and inference infrastructure: GPU pipelines, streaming inference, quantization, and compression.
- Research on continual learning, catastrophic forgetting, self-supervised objectives for tables, and post-training quantization.
Language models, retrieval, and geospatial intelligence
- Parameter-efficient fine-tuning and domain instruction tuning of language models.
- Retrieval over millions of records with hybrid search, vector databases, and cross-encoder reranking.
- ML pipelines for classification, prediction, and spatial features, with a GIS and 3D layer for building geometry, setbacks, floor-area ratio, and density.
- GPU inference, AI-aware caching, fault-tolerant retrieval, and auto-scaling on AWS.
Industrial AI and digital twins
- Predictive maintenance with deep learning on IoT telemetry: vibration, thermal profiling, and remaining useful life.
- Computer-vision quality control with real-time rejection.
- Digital twins backed by physics-informed networks, surrogates, and Bayesian optimization.
- Anomaly detection on high-frequency sensor streams with real-time alerting and root-cause analysis.
- Reinforcement-learning agents for scheduling, resource allocation, and energy optimization.
- LLM assistants and semantic search over maintenance logs, plus big-data pipelines for OEE, forecasting, yield, and energy analytics.
Reactive systems, NLP, and network intelligence
- Reactive Scala and Akka microservices for fault-tolerant, high-throughput workloads.
- NLP chatbots with intent classification and sentiment analysis, plus vector-database retrieval.
- Digital twins and neural simulation for predictive maintenance and capacity planning.
- Deep learning for image, speech, and video analytics, and reinforcement learning for traffic optimization.
Machine learning on commerce platforms
Machine learning for recommendations, search relevance, dynamic pricing, demand forecasting, and booking-trend prediction on high-traffic platforms, with real-time integrations, ETL pipelines, CI/CD, caching, load balancing, and BI dashboards.
Foundations: web systems and security
Scalable web systems built on ASP.NET, SOAP, and XML with partner integrations, encryption and secure authentication for user and payment data, and agile project methods.
Focus
Data Language Models and foundation models for structured data, real-time and streaming AI, continual learning, self-supervised objectives for tables, advanced attention, post-training quantization, efficient inference, and agentic systems. My doctoral research is on neural networks, spiking neural networks, and neuromorphic computing.
AI and machine learning
The methods behind the models I design, train, and evaluate.
LLMs, retrieval, and evaluation
Current models, how I adapt them, and how I measure and protect them.
Agentic AI
Systems where models plan, call tools, and act over many steps.
Foundation models and applied ML
Building models from scratch, tabular models, and the classical tasks around them.
Data engineering and lakehouse
How I build governed, streaming, and protected data platforms on AWS and open table formats.
How I work with data
- Cloud analytics. I work across the AWS data analytics stack: Amazon S3, the Glue Data Catalog, Lake Formation, EMR (on EKS and Serverless), Athena, DataSync, and Amazon MSK.
- Open table formats. Deep hands-on experience with Apache Iceberg: cataloging, compaction, and schema evolution.
- Processing. Apache Spark and PySpark, with Spark Streaming for batch, micro-batch, and real-time workloads.
- Medallion lakehouse. I design and implement Bronze, Silver, and Gold layers end to end.
- Tokenization and encryption at scale. Protegrity, Thales, format-preserving encryption, and Spark-UDF de-identification pipelines.
- Event-driven architecture. Kafka and MSK, Kafka Connect with single message transforms, Schema Registry, and MirrorMaker 2 or MSK Replicator.
- Relational databases and sync. Amazon RDS, PostgreSQL, and Oracle, with data synchronization techniques.
- DataOps. CI/CD with GitLab, infrastructure as code with Terraform and AWS CDK, schema-drift detection, and data contract validation.
- Governance. Enterprise data catalogs and governance tools, including Informatica Axon and EDC, and AWS Lake Formation.
Engineering and infrastructure
Languages, services, serving, cloud, and the operational layer.
Domains
Industries and problem areas I have worked in.
Education and languages
- Integrated doctorate in machine learning and deep learning, Altinbas and Beykent Universities (expected fall 2026).
- B.Sc. in Computer Engineering, Eastern Mediterranean University (2003).
- Certifications: PMI, PSM (Scrum), and Microsoft MCITP, MCPD, MCP, and MCTS tracks.
- Languages: Turkish (native), English (fluent).
Elsewhere
Find me on LinkedIn, or read the paper on arXiv. The articles explain ideas in plain language, and the projects are things I have built.
Contact
Write to ozer@ozercemkelahmet.com.
Key facts
What is a large tabular foundation model?
A large transformer pre-trained on a very wide range of tables, so it can predict, fill in missing values, flag anomalies, and match records on a new table in one pass, without training a new model for each dataset.
What is a Data Language Model (DLM)?
A Data Language Model is a foundation model that reads raw tables directly, the way a language model reads text, with no serialization, no preprocessing, and no column names. The category is defined in the arXiv paper 2605.06290 by Özer Cem KELAHMET.
What is Schema-2?
Schema-2 is the large tabular foundation model built by Özer Cem KELAHMET. It has more than 1.5 billion parameters, uses axial attention over rows and columns, a Perceiver-style global path, and a schema path, and one checkpoint of about 2.18 billion parameters carries every capability.
What can the model do?
Row-level prediction, missing-value imputation, anomaly detection, probabilistic regression, continual learning across tasks, entity and schema matching, blind sector identification, and prediction that does not depend on column names.
Who is Özer Cem KELAHMET?
A technology and AI leader who builds Data Language Models, large tabular foundation models, and real-time AI systems, end to end from research and architecture to production.
Where can I read the paper?
The paper is on arXiv: https://arxiv.org/abs/2605.06290. The documentation on this site covers the architecture, pre-training, evaluation, and every benchmark.