Alex is available for hire
Hire AlexAlex Burlacu
Verified Expert in Engineering
Python 3 Developer
Chișinău, Chisinau, Moldova
Toptal member since May 24, 2022
Bio
Over the years, as an experienced machine learning engineer, Alex dealt with diverse problems, ranging from computer vision to natural language processing and time series forecasting. He has worked as a single engineer on the project several times, and despite scarce data and few computational resources, he succeeded where others failed. He has acted as a machine learning team lead for the past few years. In his spare time, Alex enjoys engaging in independent lecturing and ML research.
Portfolio
Self-employed
Amazon EC2, Terraform, Ansible, Mistral AI, OpenAI, Azure...
Dreamfold
Python, Machine Learning, ProteinML, AlphaFold, BoltzGen, ProteinMPNN, Ray...
FleetRover Inc
Python, Machine Learning, Artificial Intelligence (AI), AWS Lambda, Llama 3...
Experience
- Python 3 - 8 years
- Machine Learning - 7 years
- Scikit-learn - 7 years
- Docker - 6 years
- Deep Learning - 6 years
- PyTorch - 5 years
- Team Mentoring - 5 years
- Machine Learning Operations (MLOps) - 5 years
Preferred Environment
Ubuntu, Python 3, Visual Studio Code (VS Code), Git, Docker, PyTorch, Neural Networks
The most amazing...
...thing I've made is an actively-learned multilingual BERT model used in document tagging to identify tender attributes and speed up document processing.
Work Experience
ML/MLOps Consultant
2022 - PRESENT
Self-employed
- Consulted for a news startup, helping them establish an MLOps culture and processes to retrain their NLP models reliably. Enabled a team of data scientists to use LLMs for news summarization, news fact-checking, and synthetic data generation.
- Handled a voice AI system for data gathering using Vapi, OpenAI, ElevenLabs, and AWS. Later, advised the engineering team on how to transition to a custom-built Voice AI platform and developed custom endpointing and interruption detection models.
- Advised a team of senior data scientists on optimizing their cloud costs through a combination of software optimizations and right-sizing, and moving to a simple AWS, Azure, and on-premise cloud setup (for both batch inference and training).
- Helped a startup right-size its GCP VMs for CNN fine-tuning. Used Terraform and Ansible to reliably provision VMs. Defined the architecture to support multi-model placement on A100 GPUs for training and used Spot VMs to improve cost efficiency.
- Implemented a multi-modal (vision and text) photo editing application using GPT-4 and DALL-E 3. The application was deployed on Google Cloud Run using Docker.
Technologies: Amazon EC2, Terraform, Ansible, Mistral AI, OpenAI, Azure, Google Cloud Platform (GCP), Cloud Run, GCP VMs, Llama 3, Convolutional Neural Networks (CNNs), Docker, Machine Learning Operations (MLOps), MLflow, Large Language Models (LLMs), Prompt Engineering, Large Language Model Operations (LLMOps), Synthetic Data Generation, Vector Databases, Amazon EC2 API, Fine-tuning, LoRa, Reinforcement Learning from Human Feedback (RLHF), OpenAI API, Solution Architecture, Speech Recognition, Evaluation
Machine Learning Engineer
2025 - 2025
Dreamfold
- Conducted extensive benchmarking and performance analysis of Jax-based models, namely AlphaFold hallucinations-based protein design, and identified optimal compute configuration for various design regimes.
- Implemented a worker node preference policy on Anyscale to use L4 GPUs when available, and roll back to T4s or A100s, as a second choice. This effort led to a 3x overall speedup for experiments, while being only 2x more expensive.
- Advised on interviewing practices, conducted technical interviews, and hired a VP of Engineering for the organization.
- Deployed multiple Anyscale Services with Ray Serve, for both binder protein filtering using BindCraft, and inverse folding with ProteinMPNN.
- Conducted a series of large HPO searches for optimal binder protein design parameters on a Ray cluster, using Optuna's batched evaluations to speed up the surrogate function discovery.
Technologies: Python, Machine Learning, ProteinML, AlphaFold, BoltzGen, ProteinMPNN, Ray, Google Cloud Platform (GCP), Docker, JAX, PyTorch, Optuna
Lead Data Scientist
2024 - 2025
FleetRover Inc
- Architected and implemented an email-based driver preferences update system that would receive an email with driver availability updates in plain English and would create and send API requests to the back-end system to propagate the changes.
- Architected and developed a system that would automatically collect emails with load recommendations from the client's inbox and convert them into a standardized JSON document to provide additional custom route recommendations for their drivers.
- Migrated both systems to Together.ai as they relied initially on Llama 3 70B with JSON mode deployed on OctoAI. The code was deployed on AWS Lambdas and triggered through SNS when emails hit a specific SES inbox.
- Improved an existing voice calling agent, using OpenAI GPT-4o and multiple advanced prompting techniques like the structured chain of thought. The calling infrastructure was on Vapi. It's used to clarify load details by directly talking to brokers.
Technologies: Python, Machine Learning, Artificial Intelligence (AI), AWS Lambda, Llama 3, AI Agents, Agentic AI, Amazon Simple Notification Service (SNS), Amazon Simple Email Service (SES), Vapi, JSON Mode, OpenAI GPT-4 API, Together.ai, Prompt Engineering, OpenAI API, Speech Recognition, Evaluation
NLP Expert
2023 - 2024
Undetectable AI
- Developed a custom self-hosted LLM solution for the project, deployed it using Terraform on AWS ECS, and optimized its serving on EC2 g5 instances using vLLM. Also developed load tests to assess, profile, and optimize the inference stack.
- Ran multiple text EDA analyses, including POS tags, semantic coherence, n-grams, and text readability statistics, to identify relevant structural and semantic patterns that we later used to adjust the core product's performance.
- Researched non-LLM-based solutions and automatic textual patterns mining from documents for text enhancement in an effort to create a new iteration of the product.
Technologies: PyTorch, Artificial Intelligence (AI), Natural Language Processing (NLP), Amazon SageMaker, OpenAI GPT-3 API, Graphics Processing Unit (GPU), Amazon EC2, Amazon Elastic Container Service (ECS), Large Language Models (LLMs), Terraform, OpenAI, Generative Artificial Intelligence (GenAI), Language Models, Amazon Web Services (AWS), Open-source LLMs, LLM Serving, Profiling, BERT, Prompt Engineering
Software Lead
2023 - 2024
ClearML
- Led the implementation of the ClearGPT project, a set of no-code tools to train and deploy self-hosted LLMs for enterprises. Actively shaped the project's design and roadmap at different stages, specifically MVP, demo, and customer POC.
- Tuned and deployed multiple LLaMA-based and FLAN-T5 models on AWS G5 instances for optimal price/performance ratio. Also worked with multi-GPU and multi-node training using HuggingFace Accelerate.
- Built tools to generate Q&A datasets from documentation pages, a RAG-aware dataset generation pipeline, and a custom trainer that oversamples the worst-performing examples to force the model to focus more on improving the performance of hard examples.
- Led the ClearML SDK team that developed features and ensured the timely release of both open-source and enterprise versions of packages. Actively involved in prioritizing and planning features for future releases.
- Contributed to community and enterprise support activities. Handled technical onboarding and actively advised clients on how best to leverage ClearML for their MLOps needs and ClearGPT for their enterprise GenAI setting.
- Debugged multiple issues related to dataset metadata management, pipelines API, distributed LLM training, and environment tracking for reproducibility.
- Acted as the go-to person for Google Cloud-related issues. Helped with the creation of custom machine images and support Spot instances. Provided solution architecture support for enterprise customers running ClearML on GCP.
Technologies: Python 3, PyTorch, ClearML, Amazon EC2, AWS Command Line Interface (CLI), Amazon S3 (AWS S3), Multi-GPU Training, Machine Learning, Large Language Models (LLMs), Meta Llama, Flan-T5, Question Generation, Q&A Bots, Generative Pre-trained Transformers (GPT), Docker, Design Patterns, Retrieval-augmented Generation (RAG), Python, Debugging, OpenAI GPT-3 API, Graphics Processing Unit (GPU), Generative Artificial Intelligence (GenAI), Language Models, Prompt Engineering, Fine-tuning, Data Synthesis, Artificial Intelligence (AI), Amazon Web Services (AWS), Open-source LLMs, Google Cloud Platform (GCP), Vector Databases, Synthetic Data Generation, LoRa, Solution Architecture, Technical Leadership
Machine Learning Team Lead
2019 - 2022
DevelopmentAid
- Used machine learning (ML) and deep learning for natural language processing (NLP) on documents to make data entry more efficient.
- Developed and produced multiple ML microservices, including one to classify and tag documents through named entity recognition using PyTorch and BERT, and another to deal with an imbalanced multi-output text classification using scikit-learn.
- Defined and wrote programs for fast data annotation and synthetic data enrichment for named entity recognition (NER). Increased the dataset size from a handful of well-annotated documents to more than a hundred.
- Guided the development of new ML models and implemented practices such as ML code review, cross-validation, and replicable experiments.
- Defined some MLOps practices mainly related to model serving using Ray Serve and experiment tracking with MLflow.
- Established an observability infrastructure to reduce the number of unreported errors and accelerated bug discovery from a few days to about 10 minutes. Used Jaeger and ELK and helped in the adoption of Prometheus and Grafana.
- Defined and documented the deployment process and reduced the time to deploy trained models to less than 10 minutes. Managed a Jenkins instance and used Jenkins pipelines for that.
- Established code reviews, periodic one-on-one meetings, explicit coding best practices, and agile processes like iteration planning, planning poker, and standup meetings, reducing feature cycle time by 5x and new bugs per iteration to 0.3.
- Led a team of three junior engineers since July 2020 in developing an automated data entry solution, developing and deploying new ML models, and handling our observability and CI infrastructures.
Technologies: Agile Software Development, Bash, Python 3, Docker, Docker Compose, PyTorch, RabbitMQ, Ray, Python, Flask, FastAPI, Jenkins, Jenkins Pipeline, Scikit-learn, Jaeger, Prometheus, Grafana, Machine Learning Operations (MLOps), Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP), Deep Learning, Transformers, MLflow, Team Mentoring, Machine Learning, Python 2, Artificial Intelligence (AI), JSON, Team Leadership, Pandas, Data Science, Hugging Face, BERT, Neural Networks, REST APIs, Fine-tuning, DevOps, Graphics Processing Unit (GPU), Language Models, Synthetic Data Generation, Technical Leadership
Research Intern
2021 - 2021
Universite Sorbonne Paris Nord
- Increased sample efficiency of deep learning algorithms, mixing techniques from self-supervised, semi-supervised, and few-shot learning applicable to images and other data sources.
- Used Google Colab notebooks to run experiments, then switched to Google Cloud Platform. Provisioned with Terraform and Ansible, creating a graphics processing unit (GPU) worker and a tracking server in a single bash command within one to two minutes.
- Used MLFlow for experiment tracking and a combination of Papermill and Optuna for hyperparameter optimization.
Technologies: Self-supervised Learning, Terraform, Google Cloud, Ansible, Few-shot Learning, MLflow, Hyperparameter Optimization, Optuna, Python 3, Python, Machine Learning, Artificial Intelligence (AI), Computer Vision, Machine Learning Operations (MLOps), PyTorch, Deep Learning, Neural Networks, Fine-tuning, DevOps, Graphics Processing Unit (GPU), Research
University Assistant
2019 - 2021
Technical University of Moldova
- Recreated and taught the network programming course and two lab projects focusing on concurrency primitives and networking protocols.
- Authored and lectured the real-time programming course and three lab projects covering message-based concurrency, including actor model and CSP, and message-oriented integration patterns and protocols like MQTT and XMPP.
- Overhauled and led the distributed systems and network programming courses and labs. Updated the real-time programming course and taught it as well.
- Covered diverse topics in the distributed systems course, such as data processing systems, distributed databases, microservice design patterns, and main problems of distributed systems, like the consensus, time, and exactly-once delivery.
- Mentored five final-year students for their semester project; two of them chose me as their bachelor thesis supervisor. Led labs for over 40 students per semester.
Technologies: Elixir, Python 3, HTTP, HTTP 2, Transmission Control Protocol (TCP), RabbitMQ, Docker, Actor Model, Erlang, Scala, Message Queues, Parallel Programming, REST, HATEOAS, Mentorship, Distributed Systems, Microservices, Python, REST APIs
Summer Intern
2018 - 2018
Cern
- Participated in the EP-SFT group as an associate partner, receiving a grant from the UK Science and Technology Facilities Council (STFC).
- Developed a project to benchmark the TMVA package against TensorFlow on event-by-event inference performance targeting multi-layered perceptrons for high-energy physics (HEP).
- Searched for the bottlenecks and future directions of optimization for the TMVA subpackage of the ROOT scientific package.
- Concluded that, for one-by-one and small batch (< 32) inference modes, TMVA is up to two orders of magnitude faster than TensorFlow 1.8, built from source with AVX512 enabled using a C++ inference API.
- Presented a poster about this work at a session at the EEML 2019 Summer School in Bucharest.
Technologies: C++, Bash, Docker, ROOT, TensorFlow, Keras, Bazel, Python 3, Python, Machine Learning, Deep Learning, Neural Networks
Machine Learning Engineer
2017 - 2018
Redox Entertainment
- Researched and developed neural networks for medical image analysis of oocytes for IVF. Created over ten bespoke neural network architectures using techniques like pre-training with autoencoders and siamese networks for self-supervised learning.
- Mentored and trained a Ph.D. intern for three months who became part of the team, also working on deep learning-related projects.
- Developed a specialized architecture for a small-sized, low-variance dataset of medical images with a performance on par with Google's AutoML Vision.
- Debugged a pre-processing data issue leaking the test set and wrongfully giving very high accuracy during evaluation. Prevented releasing the broken model, thus saving the company's reputation.
Technologies: Bash, TensorFlow, Docker, Keras, Scikit-image, Computer Vision, Self-supervised Learning, Deep Learning, Machine Learning, Medical Imaging, Team Mentoring, Python 3, Python, OpenCV, Artificial Intelligence (AI), Image Processing, SQL, Data Science, Neural Networks, Fine-tuning, Graphics Processing Unit (GPU), Research, Google Cloud Machine Learning
Co-founder and CTO
2017 - 2017
BookVoyager
- Developed a search and content-based recommendation system for fiction books that extracts features from raw text and provides recommendations based on those features.
- Implemented logging for faster troubleshooting and defined the architecture as a multiservice system.
- Built the feature extraction and recommendation sub-systems based on token-level and whole-text analysis with SpaCy.
- Participated in customer interviews, defined both business and development processes, and pitched the project at various venues.
- Sped up the computation of recommendation results 85x by using a pre-allocated array and used profiling to identify the bottleneck.
Technologies: Python 3, Python, Unit Testing, SpaCy, Scikit-learn, MongoDB, Flask, REST, XML-RPC, Git, Sentiment Analysis, Natural Language Processing (NLP), Generative Pre-trained Transformers (GPT), JSON, REST APIs
Experience
Serverless Platform
I designed and implemented a serverless platform from scratch, an AWS Lambda using Python, FastAPI, and Docker. With Vue.js, I wrote a rudimentary UI for it. The platform allowed the creation and running of arbitrary Python code within Docker containers.
To enrich its functionality, I added a few other services like RabbitMQ, Minio, PostgresSQL, MongoDB, and Apache Tika. To make it easier to use, I wrote an API Gateway-like service, a TCP server translating HTTP requests into messages and sending the responses back to the caller as HTTP responses.
The project later became the base of an independently taught course on distributed systems design. It was a free course, with 25 students enrolled, 11 of which received certificates of completion.
To enrich its functionality, I added a few other services like RabbitMQ, Minio, PostgresSQL, MongoDB, and Apache Tika. To make it easier to use, I wrote an API Gateway-like service, a TCP server translating HTTP requests into messages and sending the responses back to the caller as HTTP responses.
The project later became the base of an independently taught course on distributed systems design. It was a free course, with 25 students enrolled, 11 of which received certificates of completion.
Alex's Occasional Blog Posts | Personal Blog
In my personal blog, hosted on GitHub, I write mainly about advanced machine learning and deep learning techniques, but also about things that I have at least some non-trivial knowledge of, like distributed systems, elixir language, concurrency and parallelism, and topics relevant for senior engineers and team leads.
I created it using Jekyll, customized some of the templates, and added Google Analytics and Google Tag Manager.
I created it using Jekyll, customized some of the templates, and added Google Analytics and Google Tag Manager.
Lightweight MLOps Template for AI Research
A set of tools to speed up the process of running and tracking multiple DL/AI experiments on cloud infrastructure, specifically tuned spot instances to lower the overall cost or iterating on models. To make the tool easy to use with Jupyter, it has custom MLFlow logging. It uses a combination of Papermill and Optuna for running a hyperparameter optimization on Jupyter Notebooks with minimal adjustments.
Moldova's National Python and AI Curriculum
https://mecc.gov.md/sites/default/files/curriculum_ia_aprobat_cnc.pdfI was the group coordinator developing the national Artificial Intelligence curriculum for high-school students. As part of the effort, we formed the structure of the course to make it engaging for high school students. I created educational materials consisting of fourteen-course notes and exercise notebooks in Romanian. As part of the course, students would learn about the main applications of AI, briefly learn about AI for computer vision and natural language processing, then learn about main machine learning algorithms, then the fundamentals of linear algebra and exploratory data analysis. Finally, various social and non-technical aspects of AI include legal and social issues around self-driving cars, deep fakes, face recognition technologies, and LLMs.
Education
2019 - 2021
Master's Degree in Computer Science
Stefan cel Mare University - Suceava, Romania
2019 - 2021
Master's Degree in Computer Science
Technical University of Moldova - Chisinau, Moldova
2015 - 2019
Bachelor's Degree in Computer Science
Technical University of Moldova - Chisinau, Moldova
Certifications
DECEMBER 2022 - DECEMBER 2024
Google Cloud Certified Professional Machine Learning Engineer
Google Cloud
DECEMBER 2022 - DECEMBER 2024
Google Cloud Certified Professional Cloud Architect
Google Cloud
JUNE 2022 - JUNE 2025
Certified Kubernetes Application Developer (CKAD)
The Cloud Native Computing Foundation (CNCF)
JUNE 2022 - PRESENT
Deep Learning Engineer
Workera
Skills
Libraries/APIs
Scikit-learn, REST APIs, PyTorch, TensorFlow, OpenAI API, Jenkins Pipeline, Pandas, Keras, Vue, OpenCV, SpaCy, Amazon EC2 API, JAX
Tools
Git, Docker Compose, RabbitMQ, Jekyll, Google Analytics, Jenkins, Grafana, Scikit-image, Terraform, Ansible, Bazel, Kustomize, Helm, BigQuery, AWS Command Line Interface (CLI), Amazon Elastic Container Service (ECS), Amazon SageMaker, Amazon Simple Notification Service (SNS), Amazon Simple Email Service (SES)
Languages
Python 3, Python, Elixir, Bash, SQL, C++, C, Python 2, Lisp, HTML, CSS, Java 8, Erlang, Scala
Paradigms
REST, Functional Programming, DevOps, Unit Testing, Object-oriented Analysis & Design (OOAD), Object-oriented Programming (OOP), Agile Software Development, Serverless Architecture, Parallel Programming, Actor Model, Microservices, Design Patterns, Synthetic Data Generation
Platforms
Docker, ClearML, Ubuntu, Amazon EC2, Amazon Web Services (AWS), Kubernetes, Jupyter Notebook, Google Cloud Platform (GCP), Visual Studio Code (VS Code), Azure, Cloud Run, AWS Lambda, Together.ai
Frameworks
Flask, Ray, Optuna
Storage
JSON, Google Cloud, MongoDB, XML-RPC, PostgreSQL, Amazon S3 (AWS S3)
Other
Deep Learning, Machine Learning, Machine Learning Operations (MLOps), Natural Language Processing (NLP), Data Science, Artificial Intelligence (AI), BERT, Neural Networks, Generative Pre-trained Transformers (GPT), Fine-tuning, Large Language Models (LLMs), Language Models, University Teaching, Team Mentoring, FastAPI, Self-supervised Learning, Computer Vision, Team Leadership, Hugging Face, Retrieval-augmented Generation (RAG), Graphics Processing Unit (GPU), Generative Artificial Intelligence (GenAI), Prompt Engineering, Data Synthesis, OpenAI, Open-source LLMs, Mistral AI, Llama 3, Cloud, Large Language Model Operations (LLMOps), Vector Databases, Agentic AI, LoRa, Reinforcement Learning from Human Feedback (RLHF), Technical Leadership, Evaluation, Distributed Systems, Cloud Computing, MinIO, Serverless, Transmission Control Protocol (TCP), HTTP, Coding, HATEOAS, Jaeger, Prometheus, Transformers, MLflow, Medical Imaging, Few-shot Learning, Hyperparameter Optimization, ROOT, HTTP 2, Message Queues, Mentorship, Image Processing, Sentiment Analysis, Data Engineering, Multi-GPU Training, Meta Llama, Flan-T5, Question Generation, Q&A Bots, OpenAI GPT-3 API, Debugging, Research, GCP VMs, Google Cloud Machine Learning, Convolutional Neural Networks (CNNs), LLM Serving, Profiling, AI Agents, Vapi, JSON Mode, OpenAI GPT-4 API, Solution Architecture, Speech Recognition, ProteinML, AlphaFold, BoltzGen, ProteinMPNN
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