Anzor is available for hire
Hire AnzorAnzor Gozalishvili
Verified Expert in Engineering
Machine Learning Developer
Tbilisi, Georgia
Toptal member since November 17, 2021
Bio
Anzor is a senior machine learning and MLOps engineer with 8+ years of experience building production ML systems, data pipelines, and cloud infrastructure. He has designed ML platforms with Dagster, W&B, MLflow, Spark, AWS, and Azure and built agentic systems using LangGraph, RAG, vector search, and MCP. His background spans NLP, document intelligence, deep learning, bioinformatics, and large-scale model training. He is also a published researcher in NLP and computational biology.
Portfolio
Genentech
Dagster, Amazon SageMaker, Machine Learning Operations (MLOps), LangChain...
KYROS Insights
PyTorch Lightning, Regression, Clustering Algorithms, Profiling, MLflow...
Pfizer - Manufacturing Operations Solutions
Python, Natural Language Processing (NLP)...
Experience
- Docker - 4 years
- PyTorch - 3 years
- SpaCy - 3 years
- Amazon EC2 - 3 years
- Amazon S3 (AWS S3) - 3 years
- Machine Learning Operations (MLOps) - 2 years
- Amazon SageMaker - 2 years
- Amazon SageMaker Pipelines - 1 year
Work Experience
Senior MLOps Engineer
2024 - 2026
Genentech
- Designed and implemented a Dagster-based MLOps platform for drug-discovery ML, adding reproducible orchestration, S3 artifact versioning, experiment lineage, and a custom W&B integration.
- Built a LangGraph-based internal agent for scientists, integrating internal databases, Amazon S3 Vectors, scientific code execution, GPT models, and reusable MCP tools.
- Supported migration of GPU training workloads to AWS ParallelCluster and Slurm, containerizing jobs with Docker, profiling PyTorch/CUDA performance, and benchmarking GPU and CPU instances.
- Partnered with bioinformaticians to translate complex NGS and potency model requirements into technical specifications, aligning R&D drug-discovery objectives with engineering execution for specialized workflows.
- Led the unification of scientific ML pipelines into a shared framework covering ingestion, normalization, feature extraction, training, evaluation, and reporting.
Technologies: Dagster, Amazon SageMaker, Machine Learning Operations (MLOps), LangChain, LangGraph, AI Agents, Model Context Protocol (MCP), AI Tools, Claude Code, Large Language Models (LLMs), Large Language Model Operations (LLMOps), Python, Retrieval-augmented Generation (RAG), Vector Search, Claude, Amazon Elastic Container Registry (ECR), Amazon Elastic Container Service (ECS), FastAPI, Amazon Bedrock AgentCore, Agentic AI, Bioinformatics, Generative Artificial Intelligence (GenAI), Artificial Intelligence (AI), RAG Pipelines, LLM Reasoning, AWS Lambda, Agentic AI Systems, Agentic Frameworks, LLM Integration, Kubernetes, Vector Databases
Senior ML Engineer
2022 - 2024
KYROS Insights
- Designed and optimized multi-target regression models in PyTorch Lightning for large tabular time-series datasets, including custom output heads, loss functions, and Gaussian mixture approaches.
- Reworked the Spark-to-GPU data path using pre-batching and memory-mapped tensors, eliminating batch-construction bottlenecks and reducing training runs from hours to minutes.
- Used PyTorch Profiler and CUDA-level runtime analysis to diagnose training bottlenecks and improve end-to-end model development performance.
- Built stakeholder-facing analysis in Plotly Dash and used MLflow to track experiments, compare model variants, and support iterative development.
Technologies: PyTorch Lightning, Regression, Clustering Algorithms, Profiling, MLflow, Databricks, Spark SQL, Spark ML, Model Evaluation, Prediction Markets, Predictive Modeling, Azure, Python, Azure Databricks, Artificial Intelligence (AI), Plotly, Hugging Face, Tokenization
NLP Expert | Data Scientist
2022 - 2022
Pfizer - Manufacturing Operations Solutions
- Led development of a document-processing pipeline combining OCR and computer vision models to classify content and normalize heterogeneous SOP files into a common web-document format.
- Designed extraction strategies for tables, charts, multi-column text, formulas, and other mixed-layout content across heterogeneous SOP documents.
- Evaluated document layout analysis and OCR approaches and integrated multiple models into a production-oriented processing workflow.
Technologies: Python, Natural Language Processing (NLP), Generative Pre-trained Transformers (GPT), Data Science, Amazon Web Services (AWS), Keras, PyTorch, Amazon SageMaker, Parsers, Optical Character Recognition (OCR), Amazon Textract, Tesseract, Bioinformatics, Artificial Intelligence (AI), Hugging Face, Tokenization
NLP Data Scientist
2022 - 2022
Holocene GmbH
- Built document-processing pipelines for shipping records that extracted structured values from OCR output and combined rule-based logic with fine-tuned NLP models.
- Developed extraction pipelines using LayoutLMv2, Amazon Textract Forms, and rule-based methods, with entity classification layered on top of OCR outputs.
- Added validation logic to reconcile extracted totals and figures against calculated values, then exposed structured results to the core platform through REST APIs.
Technologies: Data Science, Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP), Python, REST APIs, Optical Character Recognition (OCR), Text Mining, Machine Learning, Authentication, APIs, Amazon Textract, Hugging Face Transformers, LayoutLMv2, Docker, Amazon S3 (AWS S3), Amazon SageMaker, Proof of Concept (POC), Artificial Intelligence (AI), Hugging Face, Tokenization
Senior Machine Learning Researcher
2021 - 2021
MaxinAI
- Reproduced and extended a bitrate-ladder prediction pipeline, building batch FFmpeg workflows, extracting frame-level features, storing processed data in S3, and training classical ML models.
- Improved bitrate-ladder prediction performance by 3% over the reproduced baseline through new video-derived features.
- Experimented with deep learning approaches for extracting video features directly while reducing inference time toward production constraints.
Technologies: PyTorch, Computer Vision, Video Encoding, FFmpeg, Scikit-image, NumPy, Research, Machine Learning, Scikit-learn, Docker, Docker Compose, Jupyter Notebook, Python, Artificial Intelligence (AI), ETL, CI/CD Pipelines, Amazon Web Services (AWS), Neural Networks
Data Scientist
2021 - 2021
Delivery Hero (Outstaffed from MaxinAI)
- Performed query optimization and migration from AWS Redshift to Google BigQuery that increased readability and established better query run time.
- Incorporated new features into models, resulting in 2% performance improvements on average.
- Worked with machine learning pipelines and Airflow DAGs, resulting in a more automated workflow with less human interaction.
Technologies: Amazon Redshift, BigQuery, SQL, Machine Learning Operations (MLOps), TensorFlow, MLflow, Apache Airflow, Amazon Web Services (AWS), Scikit-learn, Data Science, Data Engineering, Agile, Amazon SageMaker, Artificial Intelligence (AI), Cloud Deployment, Big Data, ETL, CI/CD Pipelines, Containerization, AWS IAM, Data Pipelines, Time Series Forecasting
Lecturer | Teaching Assistant
2019 - 2020
MaxinAI
- Delivered lectures and hands-on workshops in machine learning and natural language processing.
- Designed practical exercises, workshop materials, and examples to support applied understanding of ML and NLP concepts.
- Supported students during technical sessions by reviewing implementations, answering questions, and helping troubleshoot modeling issues.
Technologies: Natural Language Processing (NLP), Generative Pre-trained Transformers (GPT), Docker, Scikit-learn, Jupyter Notebook, Machine Learning, Deep Learning, SpaCy, Pandas, NumPy, Workshop Facilitation, Lecturing, Statistics, Explainable Artificial Intelligence (XAI), Hugging Face, Generative Artificial Intelligence (GenAI), Artificial Intelligence (AI), Embeddings from Language Models (ELMo), Time Series Forecasting, Tokenization
Lead Machine Learning and NLP Engineer
2018 - 2020
MaxinAI
- Headed MaxinAI’s NLP chapter, working hands-on with a team of four ML engineers across multiple client projects.
- Owned the NLP and OCR pipeline for a Swiss food-regulatory product, including OCR benchmarking, table extraction, text analysis, and structured value extraction from food labels.
- Built a startup data collection and matching pipeline using web crawling, vector search, keyword, fuzzy, and semantic matching for job-candidate recommendations.
- Built a document intelligence product combining multi-vendor OCR, BERT-based question answering, normalization logic, and a feedback loop that converted user corrections into training data.
- Developed deep learning models for merchant-name extraction from bank transactions using LSTM-CRF and CNN architectures, achieving 97% F1.
- Built a semantic search engine for US legal cases using BERT embeddings, keyword extraction, and fast vector search to retrieve relevant passages.
- Built a hybrid food recommender for MenuMend using content-based ranking, collaborative filtering, menu entity extraction, keyword search, and FAISS vector retrieval, with strategies for popularity and cold-start cases.
Technologies: Docker, Flask, PyTorch, SpaCy, Pandas, Scikit-learn, Amazon EC2, Amazon S3 (AWS S3), RabbitMQ, ABBYY, Amazon Rekognition, Optical Character Recognition (OCR), Docker Compose, Gensim, TensorFlow, Machine Learning Operations (MLOps), Matplotlib, Seaborn, Scikit-image, SciPy, NumPy, PyCharm, Deep Learning, Natural Language Toolkit (NLTK), Datadog, Flask-RESTful, fastText, OpenCV, PIL, SQL, Jupyter Notebook, Word2Vec, BERT, LSTM, Sentiment Analysis, Explainable Artificial Intelligence (XAI), MongoDB, Elasticsearch, Data Science, Data Engineering, Agile, Statistics, Bayesian Statistics, Artificial Intelligence (AI), Cloud Deployment, Google Vision API, Entity Extraction, Data Analysis, Algorithms, GitHub, Amazon Web Services (AWS), XGBoost, Classification, Regression, AutoML, Containerization, AWS IAM, Terraform, Tesseract, REST APIs, Elastic, ELK (Elastic Stack), Statistical Data Analysis, Time Series Analysis, Time Series, Neural Networks, Recurrent Neural Networks (RNNs), Data Pipelines, PostgreSQL, Data Analytics, Algorithmic Trading, Recommendation Systems, Information Retrieval, PyMongo, Text Detection, Keras, Active Learning, Variational Autoencoders (VAEs), Hugging Face, Generative Artificial Intelligence (GenAI), MLflow, Neo4j, Graph Databases, Knowledge Graphs, Embeddings from Language Models (ELMo), Time Series Forecasting, Tokenization
Project History
Unified Scientific ML Platform for Drug Discovery
Problem:
5+ teams maintained hundreds of data/model pipelines across separate repos, with inconsistent schemas and mixed stacks including R, scikit-learn, and PyTorch Lightning.
Jobs were often manual, with fragmented versioning, evaluation, and deployment workflows.
Solution:
Co-designed a shared platform with 2 engineers using Dagster, PyTorch Lightning, and Hydra.
Migrated R workflows to Python, wrapped scikit-learn models into the common training framework, and standardized configs and schemas.
Built automated lineage/versioning across Dagster and W&B, with S3-backed artifacts.
Added evaluation-gated deployment to Kubernetes/FastAPI for real-time inference and Dagster jobs for batch inference.
Built a Streamlit UI for data, model, prediction, and evaluation comparison.
Outcome:
2–3× faster model onboarding/development.
5× faster cross-team result analysis.
3× less human effort required to operate and maintain workflows.
5+ teams maintained hundreds of data/model pipelines across separate repos, with inconsistent schemas and mixed stacks including R, scikit-learn, and PyTorch Lightning.
Jobs were often manual, with fragmented versioning, evaluation, and deployment workflows.
Solution:
Co-designed a shared platform with 2 engineers using Dagster, PyTorch Lightning, and Hydra.
Migrated R workflows to Python, wrapped scikit-learn models into the common training framework, and standardized configs and schemas.
Built automated lineage/versioning across Dagster and W&B, with S3-backed artifacts.
Added evaluation-gated deployment to Kubernetes/FastAPI for real-time inference and Dagster jobs for batch inference.
Built a Streamlit UI for data, model, prediction, and evaluation comparison.
Outcome:
2–3× faster model onboarding/development.
5× faster cross-team result analysis.
3× less human effort required to operate and maintain workflows.
Scientific AI Agent for Knowledge & Workflow Access
Problem:
Scientists worked across many tools, databases, notebooks, models, and experiment outputs, making knowledge and workflows fragmented.
Reusing findings, analysis utilities, and deployed models required switching between multiple systems and interfaces.
Solution:
Built a LangGraph-based agent with a ChatGPT-style OpenWebUI interface.
Connected internal databases, S3 data, S3 Vector Search, shared memories, Jupyter sandbox execution, and custom MCP tools.
Enabled scientists to retrieve indexed findings, run code, generate plots, perform analysis, and invoke deployed models through natural language.
Exposed reusable scientific utilities and inference APIs as shared MCP tools.
Outcome:
Created a single interface for shared knowledge, analysis, and model access across tens of teams and hundreds of scientists.
Reduced tool switching and made access to scientific resources and prior findings dramatically faster, with workflows estimated at up to 10× faster.
Scientists worked across many tools, databases, notebooks, models, and experiment outputs, making knowledge and workflows fragmented.
Reusing findings, analysis utilities, and deployed models required switching between multiple systems and interfaces.
Solution:
Built a LangGraph-based agent with a ChatGPT-style OpenWebUI interface.
Connected internal databases, S3 data, S3 Vector Search, shared memories, Jupyter sandbox execution, and custom MCP tools.
Enabled scientists to retrieve indexed findings, run code, generate plots, perform analysis, and invoke deployed models through natural language.
Exposed reusable scientific utilities and inference APIs as shared MCP tools.
Outcome:
Created a single interface for shared knowledge, analysis, and model access across tens of teams and hundreds of scientists.
Reduced tool switching and made access to scientific resources and prior findings dramatically faster, with workflows estimated at up to 10× faster.
SageMaker ML Platform for Scientific Workflows
Problem:
50+ scientists were running ML and bioinformatics workflows in local/HPC environments with project-specific dependencies, local paths, and setup friction.
Moving experiments into production required significant rework, and many users were unfamiliar with the SageMaker SDK.
Solution:
Productionized workflows into hierarchical Docker images with shared and project-specific layers.
Built a YAML-based interface for defining pipeline steps without direct SageMaker SDK usage.
Enabled the same environments for notebook prototyping and pipelines covering processing, training, evaluation, real-time inference, and batch inference.
Partnered with DevOps on automated image builds and ECR publishing.
Outcome:
Improved reproducibility and reduced environment conflicts.
Made experimentation 5× faster.
Reduced prototype-to-production effort by 10×.
50+ scientists were running ML and bioinformatics workflows in local/HPC environments with project-specific dependencies, local paths, and setup friction.
Moving experiments into production required significant rework, and many users were unfamiliar with the SageMaker SDK.
Solution:
Productionized workflows into hierarchical Docker images with shared and project-specific layers.
Built a YAML-based interface for defining pipeline steps without direct SageMaker SDK usage.
Enabled the same environments for notebook prototyping and pipelines covering processing, training, evaluation, real-time inference, and batch inference.
Partnered with DevOps on automated image builds and ECR publishing.
Outcome:
Improved reproducibility and reduced environment conflicts.
Made experimentation 5× faster.
Reduced prototype-to-production effort by 10×.
Shipping Document Intelligence & Data Extraction
https://www.holocene.eu/supplier-reliabilityProblem:
Shipping documents arrived in many formats and layouts, making reliable classification and structured field extraction difficult across document types such as bills of lading, invoices, packing lists, and letters of credit.
Solution:
Built document and bundle classification pipelines using OCR text and lightweight ML classifiers.
Combined AWS Textract, UBIAI annotations, rules, LayoutLMv2, Detectron-based CV, and NLP models for document-specific extraction.
Developed extraction flows for forms, tables, entities, container numbers, shipping labels, signatures, and related fields.
Exposed the pipelines through REST APIs with evaluation, unit testing, load testing, and containerized deployment.
Outcome:
Built a production document-intelligence pipeline tuned for high-precision extraction, balancing conservative predictions with ~84% F1 overall across entity extraction tasks.
Shipping documents arrived in many formats and layouts, making reliable classification and structured field extraction difficult across document types such as bills of lading, invoices, packing lists, and letters of credit.
Solution:
Built document and bundle classification pipelines using OCR text and lightweight ML classifiers.
Combined AWS Textract, UBIAI annotations, rules, LayoutLMv2, Detectron-based CV, and NLP models for document-specific extraction.
Developed extraction flows for forms, tables, entities, container numbers, shipping labels, signatures, and related fields.
Exposed the pipelines through REST APIs with evaluation, unit testing, load testing, and containerized deployment.
Outcome:
Built a production document-intelligence pipeline tuned for high-precision extraction, balancing conservative predictions with ~84% F1 overall across entity extraction tasks.
SOP Document Intelligence & Web Normalization
Problem:
Pfizer SOPs were mostly PDFs with highly inconsistent formatting, including multi-column text, tables, charts, headers, and other mixed layouts. The goal was to migrate them into a single web interface with a consistent document structure.
Solution:
Led a 3-person team building a document-processing pipeline combining multiple OCR engines, including AWS Textract and Google OCR.
Combined OCR outputs to improve text recovery across different document formats and scan qualities.
Trained Detectron-style computer vision models to detect and classify layout elements such as text regions, tables, charts, and other structural components.
Reconstructed detected content into a normalized document representation suitable for the web platform.
Outcome:
Created a unified processing pipeline that transformed heterogeneous SOP PDFs into consistently structured web documents, enabling documents with very different layouts to be presented and processed through the same interface.
Pfizer SOPs were mostly PDFs with highly inconsistent formatting, including multi-column text, tables, charts, headers, and other mixed layouts. The goal was to migrate them into a single web interface with a consistent document structure.
Solution:
Led a 3-person team building a document-processing pipeline combining multiple OCR engines, including AWS Textract and Google OCR.
Combined OCR outputs to improve text recovery across different document formats and scan qualities.
Trained Detectron-style computer vision models to detect and classify layout elements such as text regions, tables, charts, and other structural components.
Reconstructed detected content into a normalized document representation suitable for the web platform.
Outcome:
Created a unified processing pipeline that transformed heterogeneous SOP PDFs into consistently structured web documents, enabling documents with very different layouts to be presented and processed through the same interface.
Actuarial Deep Learning & Training Optimization
https://www.kyros.com/solutions/proving-loyalty-program-roi/Intro:
I worked with Kyros Insights leadership on deep learning models for large actuarial time-series datasets.
Problem:
Targets were heterogeneous and zero-inflated, with different skewed non-zero distributions. Modeling them jointly created complex multimodal behavior. Shallow MLP could use very large batches, making batch construction and Spark-to-GPU data movement the main training bottleneck.
My Contributions:
I designed multi-target PyTorch Lightning models around actuarial target groups defined through domain expertise. I experimented with Gaussian-mixture and distribution-parameter outputs and built non-negative regression heads that clamped negative predictions while preserving gradient flow near zero. I also redesigned the data path using pre-batching and memory-mapped tensors.
Outcome:
The modeling approach handled zero-inflated targets better than standard regression, while the optimized pipeline reduced training runs from hours to minutes.
I worked with Kyros Insights leadership on deep learning models for large actuarial time-series datasets.
Problem:
Targets were heterogeneous and zero-inflated, with different skewed non-zero distributions. Modeling them jointly created complex multimodal behavior. Shallow MLP could use very large batches, making batch construction and Spark-to-GPU data movement the main training bottleneck.
My Contributions:
I designed multi-target PyTorch Lightning models around actuarial target groups defined through domain expertise. I experimented with Gaussian-mixture and distribution-parameter outputs and built non-negative regression heads that clamped negative predictions while preserving gradient flow near zero. I also redesigned the data path using pre-batching and memory-mapped tensors.
Outcome:
The modeling approach handled zero-inflated targets better than standard regression, while the optimized pipeline reduced training runs from hours to minutes.
Food Label Intelligence & Regulatory Compliance
https://www.digicomply.com/Problem:
Food labels came from varied angles and layouts, with mixed fonts, tables, rotated text, and complex nutrition sections. The system needed to extract reliable structured data so validators could review labels for regulatory compliance and potential violations.
Solution:
Led the NLP workstream in a 0→1 product built by two teams.
Benchmarked Google OCR, AWS Textract, and other providers, then normalized their outputs.
Matched OCR text to CV-detected label regions, corrected region classes using textual signals, and reconstructed reading order for complex layouts.
Built hybrid extraction using rules, spaCy NER, scikit-learn, and lightweight BERT models.
Used RabbitMQ for asynchronous processing and routed results to a validation UI.
Outcome:
Delivered an end-to-end pipeline with 90%+ precision and recall, helping food validators review labels faster, extract nutrition and allergen data, and identify potential compliance issues with less manual effort.
Food labels came from varied angles and layouts, with mixed fonts, tables, rotated text, and complex nutrition sections. The system needed to extract reliable structured data so validators could review labels for regulatory compliance and potential violations.
Solution:
Led the NLP workstream in a 0→1 product built by two teams.
Benchmarked Google OCR, AWS Textract, and other providers, then normalized their outputs.
Matched OCR text to CV-detected label regions, corrected region classes using textual signals, and reconstructed reading order for complex layouts.
Built hybrid extraction using rules, spaCy NER, scikit-learn, and lightweight BERT models.
Used RabbitMQ for asynchronous processing and routed results to a validation UI.
Outcome:
Delivered an end-to-end pipeline with 90%+ precision and recall, helping food validators review labels faster, extract nutrition and allergen data, and identify potential compliance issues with less manual effort.
Data Pipeline Optimization and Migration
Brief
I worked on a customer-retention ML pipeline in which data preparation, feature engineering, and orchestration were distributed across multiple systems.
Problem
The existing workflow included a recurring BigQuery-to-Redshift transfer before downstream SQL processing, adding unnecessary data movement and operational overhead. Parts of the ML workflow also relied on manual execution.
My Contributions
I refactored the pipeline so downstream data preparation ran directly in BigQuery, removing the weekly transfer to Redshift. I also engineered new features for the time-series retention model and migrated existing ML workflows into Airflow DAGs for more reliable, automated execution.
Outcome
The changes simplified the data flow, reduced manual pipeline operations, and improved the retention model by approximately 2% through additional feature engineering.
I worked on a customer-retention ML pipeline in which data preparation, feature engineering, and orchestration were distributed across multiple systems.
Problem
The existing workflow included a recurring BigQuery-to-Redshift transfer before downstream SQL processing, adding unnecessary data movement and operational overhead. Parts of the ML workflow also relied on manual execution.
My Contributions
I refactored the pipeline so downstream data preparation ran directly in BigQuery, removing the weekly transfer to Redshift. I also engineered new features for the time-series retention model and migrated existing ML workflows into Airflow DAGs for more reliable, automated execution.
Outcome
The changes simplified the data flow, reduced manual pipeline operations, and improved the retention model by approximately 2% through additional feature engineering.
QA-based Document Intelligence for Property Management
https://loftyworks.com/ukProduct Context
ExtractHD was built for Rentancy Ltd to help property managers extract structured information from rental documents and was later acquired by Loftyworks in the UK.
Problem
Extracting fields across dozens of document categories normally required substantial labeled data and separate NER models.
Solution
Built an SDK to normalize OCR outputs and preserve word-level links between documents and the UI.
Used BERT-style QA models for paragraph-level extraction, asking entity-specific questions and ranking candidate answers.
Supported tens of document categories with limited initial labeled data.
Converted property managers’ UI corrections into training examples for periodic retraining.
Outcome
Created a self-improving extraction workflow that reduced upfront annotation needs and continuously improved from real user feedback.
ExtractHD was built for Rentancy Ltd to help property managers extract structured information from rental documents and was later acquired by Loftyworks in the UK.
Problem
Extracting fields across dozens of document categories normally required substantial labeled data and separate NER models.
Solution
Built an SDK to normalize OCR outputs and preserve word-level links between documents and the UI.
Used BERT-style QA models for paragraph-level extraction, asking entity-specific questions and ranking candidate answers.
Supported tens of document categories with limited initial labeled data.
Converted property managers’ UI corrections into training examples for periodic retraining.
Outcome
Created a self-improving extraction workflow that reduced upfront annotation needs and continuously improved from real user feedback.
Content-adaptive Video Quality & Bitrate Prediction
https://cloud.qencode.com/per-title-encodingProblem
Efficient video encoding depends on content: simple scenes can tolerate stronger compression, while complex or high-motion scenes require more bitrate to preserve quality. Existing approaches relied on expensive quality evaluation across encoding configurations.
Solution
Reproduced the existing bitrate-ladder prediction research pipeline and established a working baseline.
Segmented videos into scenes and sampled representative frames from each segment.
Introduced new content features, including JPEG complexity as a proxy for frame compressibility, alongside motion/scene characteristics.
Trained ML models to predict expected visual quality for candidate encoding settings and used those predictions to guide parameter selection.
Outcome
Improved prediction performance by ~3% over the reproduced baseline, enabling more content-aware encoding decisions while reducing the need for exhaustive quality evaluation across candidate encodes.
Efficient video encoding depends on content: simple scenes can tolerate stronger compression, while complex or high-motion scenes require more bitrate to preserve quality. Existing approaches relied on expensive quality evaluation across encoding configurations.
Solution
Reproduced the existing bitrate-ladder prediction research pipeline and established a working baseline.
Segmented videos into scenes and sampled representative frames from each segment.
Introduced new content features, including JPEG complexity as a proxy for frame compressibility, alongside motion/scene characteristics.
Trained ML models to predict expected visual quality for candidate encoding settings and used those predictions to guide parameter selection.
Outcome
Improved prediction performance by ~3% over the reproduced baseline, enabling more content-aware encoding decisions while reducing the need for exhaustive quality evaluation across candidate encodes.
Georgian LLM Corpus (ACL Datasets)
I developed a large-scale Georgian corpus for LLM training, filling a critical resource gap for low-resource languages. I managed complex data collection and cleaning pipelines and reviewed the ACL ARR 2024 dataset chapter, but it was unpublished due to compliance with local regulations on digital data ownership.
Syntactic Annotation of Georgian in the UD Schemes (Springer Nature)
https://link.springer.com/chapter/10.1007/978-3-032-14339-6_6Problem
Georgian had limited syntactically annotated resources suitable for modern NLP. Building a useful treebank required not only annotation but also careful sampling to ensure the dataset represented diverse topics and sentence complexity levels.
Contribution
Helped define the research direction and dataset construction strategy.
Collected and sampled Georgian text from sources such as Wikipedia using unsupervised topic categorization, clustering, and sentence-complexity-based selection.
Worked with linguists who annotated the selected sentences using an existing annotation scheme plus manual corrections.
Trained Universal Dependencies pipeline models on the resulting dataset and evaluated the quality of the treebank for downstream parsing.
Outcome
Contributed to a 3,013-sentence Georgian UD treebank and reusable syntactic resources for low-resource NLP. I presented the work at TLT 2024 in Hamburg (https://aclanthology.org/2024.tlt-1.5/) and also published it as a chapter in a Springer book.
Georgian had limited syntactically annotated resources suitable for modern NLP. Building a useful treebank required not only annotation but also careful sampling to ensure the dataset represented diverse topics and sentence complexity levels.
Contribution
Helped define the research direction and dataset construction strategy.
Collected and sampled Georgian text from sources such as Wikipedia using unsupervised topic categorization, clustering, and sentence-complexity-based selection.
Worked with linguists who annotated the selected sentences using an existing annotation scheme plus manual corrections.
Trained Universal Dependencies pipeline models on the resulting dataset and evaluated the quality of the treebank for downstream parsing.
Outcome
Contributed to a 3,013-sentence Georgian UD treebank and reusable syntactic resources for low-resource NLP. I presented the work at TLT 2024 in Hamburg (https://aclanthology.org/2024.tlt-1.5/) and also published it as a chapter in a Springer book.
Comparative Analysis of Genetic Perturbation Models
Benchmarked deep learning models including scGPT and CPA against classical and linear baselines for genetic perturbation prediction. Evaluated generalization across public datasets and compared accuracy, complexity, and computational cost.
Forensic Eye Color Prediction with IrisPlex
Problem
Forensic DNA models such as IrisPlex predict blue and brown eyes well, but intermediate eye colors are difficult to classify and depend heavily on how eye-color phenotypes are defined.
Contribution
Analyzed high-resolution iris images and quantitative CIELAB color representations from 238 individuals.
Evaluated clustering-based definitions of eye color as an alternative to subjective visual categories.
Compared the resulting phenotype groups using the IrisPlex genetic prediction model.
Outcome
The clustering-based phenotype definition significantly improved prediction accuracy, particularly for brown eyes. The work was published in Scientific Reports in 2022, where I was listed as an equal co-author.
Forensic DNA models such as IrisPlex predict blue and brown eyes well, but intermediate eye colors are difficult to classify and depend heavily on how eye-color phenotypes are defined.
Contribution
Analyzed high-resolution iris images and quantitative CIELAB color representations from 238 individuals.
Evaluated clustering-based definitions of eye color as an alternative to subjective visual categories.
Compared the resulting phenotype groups using the IrisPlex genetic prediction model.
Outcome
The clustering-based phenotype definition significantly improved prediction accuracy, particularly for brown eyes. The work was published in Scientific Reports in 2022, where I was listed as an equal co-author.
IOAI AI Training Program — NLP Lecturer
https://www.youtube.com/watch?v=XNxX7xQZAuM&list=PL5PFQITWvGrdssEHIIe4CS7CWHNnSEXp9Contributed to Georgia’s national AI Olympiad preparation program by delivering three lectures for 10th–12th grade students on recurrent neural networks, sequence modeling, and Transformers. The sessions combined conceptual explanations with practical machine learning examples and were recorded and published online. The broader training program prepared students for international AI competitions. Georgia’s national team later earned two gold, three silver, and two bronze medals at IOAI 2026, finishing 7th among 109 participating countries. I contributed specifically as an NLP lecturer rather than as the team coach or program lead.
Machine Learning and NLP Educator
https://github.com/MaxinAI/school-of-aiIntro
As one of three lecturers at MaxinAI’s School of AI, I delivered lectures and hands-on workshops in Natural Language Processing and classical Machine Learning.
Focus
The goal was to help students connect core ML and NLP concepts to practical implementations and real-world modeling workflows.
My Contributions
I prepared and delivered technical sessions, explained classical ML and NLP methods, and guided students through practical exercises and projects. The role required turning complex topics into clear, structured material for learners with different levels of experience.
Outcome
Students gained practical exposure to ML and NLP workflows alongside the underlying theory, with hands-on work reinforcing the concepts covered in lectures.
As one of three lecturers at MaxinAI’s School of AI, I delivered lectures and hands-on workshops in Natural Language Processing and classical Machine Learning.
Focus
The goal was to help students connect core ML and NLP concepts to practical implementations and real-world modeling workflows.
My Contributions
I prepared and delivered technical sessions, explained classical ML and NLP methods, and guided students through practical exercises and projects. The role required turning complex topics into clear, structured material for learners with different levels of experience.
Outcome
Students gained practical exposure to ML and NLP workflows alongside the underlying theory, with hands-on work reinforcing the concepts covered in lectures.
Education
2018 - 2020
Master's Degree in Computer Science
Tbilisi State University - Tbilisi, Georgia
2017 - 2018
Erasmus Exchange and Scholarship in Computer Science
Universitat Politecnica de Valencia (UPV) - Valencia, Span
2014 - 2018
Bachelor's Degree in Computer Science
Tbilisi State University - Tbilisi, Georgia
Certifications
JUNE 2026 - PRESENT
Building AI Agents and Agentic Workflows
IBM
JUNE 2026 - PRESENT
Getting Started with AWS Generative AI for Developers
Coursera
JULY 2021 - PRESENT
Machine Learning Modeling Pipelines in Production
Coursera
JUNE 2021 - PRESENT
Machine Learning Data Lifecycle in Production
Coursera
MARCH 2018 - PRESENT
Deep Learning Specialization
Coursera
DECEMBER 2016 - PRESENT
Machine Learning
Coursera
Skills
Libraries/APIs
Pandas, Scikit-learn, NumPy, PyTorch, SpaCy, Matplotlib, Natural Language Toolkit (NLTK), TensorFlow, OpenCV, SciPy, FFmpeg, PIL, Keras, Flask-RESTful, Amazon Rekognition, PyMongo, LSTM, Google Vision API, XGBoost, REST APIs, Hugging Face Transformers, PyTorch Lightning, Spark ML, Stanford NLP
Tools
PyCharm, Docker Compose, Gensim, Amazon SageMaker, AutoML, RabbitMQ, BigQuery, Apache Airflow, Scikit-image, Seaborn, ABBYY, Amazon Textract, Pytest, GitHub, AWS IAM, Terraform, Elastic, ELK (Elastic Stack), Spark SQL, Claude Code, Claude, Amazon Elastic Container Registry (ECR), Amazon Elastic Container Service (ECS), Plotly
Languages
Python, SQL, Markdown
Platforms
Jupyter Notebook, Docker, Linux, Amazon EC2, Amazon Web Services (AWS), Databricks, Azure, AWS Lambda, CrewAI, Kubernetes, NVIDIA CUDA, Weights & Biases
Frameworks
Flask, LangGraph, Selenium, Scrapy, AutoGen, Agentic Frameworks, Spark, Streamlit
Paradigms
Agile, Testing, ETL, Model Context Protocol (MCP)
Storage
Amazon S3 (AWS S3), Datadog, Elasticsearch, MongoDB, Cloud Deployment, Data Pipelines, PostgreSQL, Neo4j, Graph Databases, Google Cloud
Industry Expertise
Bioinformatics
Other
Natural Language Processing (NLP), Machine Learning, Word2Vec, Data Science, Generative Pre-trained Transformers (GPT), MLflow, Deep Learning, Feature Engineering, Machine Learning Operations (MLOps), Sentiment Analysis, Explainable Artificial Intelligence (XAI), Optical Character Recognition (OCR), Text Detection, Statistics, Artificial Intelligence (AI), Classification, Regression, Hugging Face, Amazon SageMaker Pipelines, Tokenization, Large Language Models (LLMs), Generative Artificial Intelligence (GenAI), Embeddings from Language Models (ELMo), Time Series Forecasting, Amazon Redshift, Pipelines, fastText, Recommendation Systems, Information Retrieval, Active Learning, Variational Autoencoders (VAEs), Time Series Analysis, Principal Component Analysis (PCA), Lecturing, Workshop Facilitation, Computer Vision, Video Encoding, Research, BERT, Data Engineering, Bayesian Statistics, Evaluation, Text Classification, Entity Extraction, Data Analysis, Algorithms, Big Data, CI/CD Pipelines, Containerization, Tesseract, Statistical Data Analysis, Time Series, Neural Networks, Recurrent Neural Networks (RNNs), Data Analytics, Algorithmic Trading, Clustering, Experimental Design, Optimization, Genomics, Biology, Clustering Algorithms, Data Visualization, Transformers, Text Mining, Authentication, APIs, LayoutLMv2, Parsers, Proof of Concept (POC), Profiling, Model Evaluation, Dagster, LangChain, AI Agents, DataTrove, Web Crawlers, Open-source LLMs, Computational Linguistics, Conference Speaking, Data Labeling, Scanpy, Prediction Markets, Predictive Modeling, AI Tools, Large Language Model Operations (LLMOps), Retrieval-augmented Generation (RAG), Vector Search, FastAPI, Agentic AI, Azure Databricks, RAG Pipelines, LLM Reasoning, Knowledge Graphs, Data Transformation, Amazon Bedrock AgentCore, BeeAi, Agentic AI Systems, LLM Integration, Vector Databases, Technical Training, WandB, Google BigQuery, Linear Algebra, Calculus, AI Evaluation, Universal Dependencies, treebanks, Tech Conferences, pytorch profiler, Robust Regression, Data Versioning
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring