Justas S.-Data & AI Engineer | GenAI / LLM | Data Platforms
Check rate
Experience
Performance Analysis Intern
Luminor Bank - CIO Office
- Built analytical and validation systems spanning 5 organisational domains, approximately 35 business areas, and 280 operational processes.
- Designed Python and SQL transformation pipelines, automated validation rules, KPI models, and reproducible analytical workflows supporting CIO-level reporting.
- Investigated data inconsistencies across source systems, transformation pipelines, metric definitions, and reporting layers using systematic root-cause analysis.
- Implemented automated checks for completeness, schema consistency, aggregation correctness, outliers, duplicated records, and metric regressions.
- Worked with business stakeholders to translate operational requirements into reusable data models, metrics, and governed analytical outputs.
- Validated analytical pipelines and reporting outputs to ensure consistency, traceability, and reliable downstream use by business stakeholders.
Principal Engineer / Solo Developer
AlphaFlow
- Architected an event-driven platform for real-time ingestion, feature computation, ML inference, deterministic replay, execution simulation, and quantitative research.
- Designed deterministic replay and regression-testing workflows for reproducing failures and validating positions, feature vectors, model outputs, and system state across repeated executions.
- Built asynchronous market-data infrastructure for BTC, ETH, SOL, and XRP with ordering guarantees, duplicate handling, recovery, synchronization, validation, telemetry, and defensive message processing.
- Developed modular, testable interfaces across ingestion, models, strategies, execution, and risk using Rust, Python, Docker, WebSockets, ONNX, and asynchronous systems.
- Built event streaming and event sourcing infrastructure with Apache Pulsar, including structured event schemas, asynchronous producers and consumers, replayable event histories, and distributed message processing.
- Designed processing guarantees around event identity, ordering, duplicate handling, idempotent consumers, lateness policies, recovery, and deterministic replay.
- Implemented reproducible event replay and regression testing using canonicalized state, per-stage hashing, and first-divergence detection across distributed processing stages.
- Improved GPU utilisation from approximately 45% to 80% by introducing batching and moving tokenisation outside the critical execution path.
- Reduced evaluation runtime from approximately 3 hours to 70 minutes through pipeline optimisation and improved execution efficiency.
- Built a FastAPI-based evaluation service supporting job submission, execution tracking, and result retrieval for approximately 2,000 evaluation runs per month.
- Worked with approximately 15 TB of logs and evaluation data while maintaining deterministic execution and debugging workflows.
Cofounder & Lead Engineer
ServeAI
- Lead development of an AI SaaS platform using Python, PyTorch, Hugging Face, PostgreSQL, vector retrieval, LLM inference, APIs, and cloud infrastructure.
- Designed constrained retrieval and generation pipelines grounding model responses exclusively in customer-specific data with deterministic retrieval, filtering, validation, and access controls.
- Built inference benchmarking workflows comparing model quality, token throughput, time-to-first-token, end-to-end latency, memory consumption, and batching strategies.
- Evaluated local and hosted inference stacks including vLLM, SGLang, llama.cpp, Hugging Face Transformers, and quantized model deployments.
- Implemented dynamic batching, streaming generation, caching, token-budget controls, structured outputs, retry logic, and inference observability.
- Own ML architecture, evaluation, deployment, backend services, data models, model integration, performance optimization, debugging, and production reliability.
- Lead backend engineering of an AI SaaS platform spanning API services, PostgreSQL, authentication, LLM systems, retrieval, payments, and Kubernetes-based cloud infrastructure.
- Design service boundaries, REST APIs, microservices, backend workflows, data models, external integrations, and AI-enabled product functionality.
- Implement GitOps-based deployment workflows with ArgoCD for declarative Kubernetes application delivery, automated synchronization, and production deployment management.
- Built structured RAG workflows grounding AI responses in customer-specific data with retrieval constraints, validation, and application-level access controls.
- Own architecture, implementation, integration testing, deployment automation, debugging, and production iteration across the full backend lifecycle.
- Designed database schemas and backend retrieval interfaces supporting structured knowledge access, semantic retrieval, and reliable AI-assisted responses.
- Designed validation and guardrail logic around model responses to improve answer reliability and ensure outputs remain grounded in authorised application data.
Software Engineer - Machine Learning Intern
OpenAI
- Built and validated Python and PyTorch workflows for model training, inference, evaluation, and large-scale ML experimentation on NVIDIA GPU infrastructure.
- Profiled an evaluation pipeline where CPU tokenization left GPUs underutilized; redesigned batching and preprocessing to increase GPU utilization from 46% to 82% and reduce runtime from 3.1 hours to 74 minutes.
- Implemented model and tensor-level correctness tests covering shapes, dtypes, masking, padding, attention outputs, numerical tolerances, deterministic execution, and cross-implementation consistency.
- Debugged fp16 and bf16 numerical instability involving NaNs, overflowing logits, unstable reductions, and gradient scaling; validated fixes against fp32 reference implementations.
- Worked with PyTorch, Hugging Face Transformers, Tokenizers, NumPy, SciPy, CUDA, Triton, NCCL, and distributed multi-GPU workloads.
- Benchmarked model implementations using latency, throughput, peak VRAM, GPU utilization, numerical error, and output-equivalence criteria.
- Built reproducible command-line evaluation workflows with pinned dependencies, deterministic seeds, containerized execution, structured configuration, and automated result validation.
- Diagnosed distributed and systems failures including NCCL synchronization issues, GPU OOMs, host-to-device bottlenecks, memory fragmentation, connection leaks, and flaky CI workloads.
- Reviewed ML implementations and AI-generated technical solutions for correctness, efficiency, reproducibility, numerical behavior, and engineering quality.
- Developed and supported production-oriented backend and machine learning systems across Python, C++, distributed services, APIs, and containerized infrastructure.
- Worked with Kubernetes, Docker, Terraform, Azure, CI systems, FastAPI, and gRPC to build, deploy, test, and operate distributed ML services.
- Built and maintained large-scale data and event-processing workflows using Kafka, Spark, Airflow, PostgreSQL, and distributed processing infrastructure.
- Contributed to service interfaces, automated testing, production debugging, reliability engineering, and performance optimization across large shared codebases.
LLM Inference Optimization Suite
- PyTorch, vLLM, SGLang, Transformers, CUDA, Triton
- Built a reproducible benchmark suite comparing Hugging Face Transformers, vLLM, SGLang, and llama.cpp across latency, token throughput, peak memory, concurrency, and output correctness.
- Tested continuous batching, KV-cache utilization, prefix caching, quantization, tensor parallelism, sequence length, and batch-size trade-offs.
- Increased throughput on representative transformer workloads by 3.4x while reducing peak GPU memory by 31% through batching, memory-layout, and inference-engine optimization.
- Added automated regression thresholds so changes failing correctness, memory, latency, or throughput requirements were rejected automatically.
Distributed Training & Failure Analysis
- PyTorch DDP/FSDP, CUDA, NCCL, Kubernetes
- Built multi-GPU training workflows using DDP and FSDP with mixed precision, gradient accumulation, checkpointing, distributed data loading, and fault-tolerant checkpoint recovery.
- Diagnosed CUDA OOMs using memory snapshots and profiler traces and distinguished parameter, activation, optimizer-state, temporary-buffer, and fragmentation-related memory usage.
- Investigated distributed hangs caused by inconsistent collective operations, dataloader failures, rank divergence, network issues, and NCCL synchronization behavior.
- Created reproducible failure cases and automated regression tests before integrating fixes into training infrastructure.
Numerical Correctness & Autograd Validation
- PyTorch, JAX, NumPy, SciPy
- Implemented transformer and neural-network components independently in PyTorch, JAX, and NumPy to compare forward outputs and gradients across frameworks.
- Diagnosed discrepancies caused by broadcasting, masking, reduction order, dtype conversion, unstable softmax calculations, initialization differences, and floating-point precision.
- Built gradient checking utilities using finite differences and high-precision reference computations for custom differentiable operations.
- Defined explicit absolute and relative error tolerances based on operation type, dtype, tensor magnitude, and accumulation depth.
Machine Learning Research Engineer
Leiden University - LIACS
- Developed research implementations of neural network architectures in PyTorch and JAX for experiments involving optimization, representation learning, sequence modeling, and numerical methods.
- Reimplemented model components directly from mathematical definitions and research papers, including attention, normalization, loss functions, optimizers, convolutional layers, and custom differentiable operations.
- Used JAX transformations including jit, grad, vmap, and pmap to build accelerated and vectorized numerical workloads.
- Verified automatic differentiation implementations using analytical derivatives, finite differences, numerical tolerance testing, and PyTorch/JAX cross-checks.
- Designed controlled experiments with fixed seeds, versioned datasets, pinned environments, configuration files, experiment tracking, and reproducible CLI entry points.
- Supervised benchmarking of training and inference implementations across CPU and GPU hardware and documented performance and numerical trade-offs.
Industry experience
See where this freelancer has spent most of their professional time.
Experienced in Education, Information Technology, and Banking and Finance.
Business area experience
See which departments and functions this freelancer has contributed to most.
Experienced in Information Technology, Research and Development, Quality Assurance, Product Development, and Business Intelligence.
Summary
Machine learning systems engineer with 8 years of experience building, training, profiling, debugging, and optimizing deep learning systems. Specialized in Python, PyTorch, JAX, numerical computing, GPU acceleration, distributed training, and high-throughput LLM inference. Experienced with vLLM, SGLang, Hugging Face Transformers, Tokenizers, CUDA, Triton, NCCL, and reproducible ML workflows. Strong track record diagnosing numerical instability, tensor shape and dtype errors, memory bottlenecks, distributed failures, and performance regressions from model code down to GPU execution. Also experienced building RAG workflows, LLM-backed applications, batch and streaming data pipelines, evaluation systems, production APIs, data platforms, and production infrastructure using Python, SQL, Spark, Kafka, Airflow, PostgreSQL, MLflow, Kubernetes, Terraform, Azure, AWS, and modern lakehouse patterns.
Skills
Programming: Python, C++, Rust, Sql, Bash, Typescript, Java, Go
Ml Frameworks: Pytorch, Jax, Flax, Hugging Face Transformers, Hugging Face Tokenizers, Scikit-Learn, Xgboost, Lightgbm
Inference: Vllm, Sglang, Llama.Cpp, Onnx Runtime, Tensorrt, Continuous Batching, Kv Caching, Quantization, Speculative Decoding
Numerical Computing: Numpy, Scipy, Tensor Algebra, Automatic Differentiation, Numerical Stability, Finite-Difference Validation, Mixed Precision
Training: Ddp, Fsdp, Gradient Accumulation, Gradient Checkpointing, Amp, Bf16/Fp16, Optimizer Tuning, Learning-Rate Scheduling
Gpu And Performance: Cuda, Triton, Nccl, Nvidia Nsight, Pytorch Profiler, Torch.Compile, Xla, Kernel Profiling, Memory Analysis
Data And Systems: Spark, Kafka, Airflow, Postgresql, Duckdb, Redis, Distributed Processing, Etl, Streaming Pipelines
Infrastructure: Kubernetes, Docker, Terraform, Aws, Azure, Linux, Fastapi, Grpc, Bazel, Github Actions, Ci/Cd
Evaluation: Reproducibility, Benchmark Design, Regression Testing, Numerical Validation, Throughput, Latency, Memory, Correctness
Model Implementation: Able To Implement Model Components Directly From Equations Rather Than Relying Exclusively On High-Level Apis.
Numerical Debugging: Experienced Diagnosing Nans, Exploding Gradients, Precision Loss, Unstable Reductions, Incorrect Broadcasting, And Dtype-Related Failures.
Performance Optimization: Measures Bottlenecks Before Optimization Using Profiling, Hardware Utilization, Throughput, Latency, And Memory Traces.
Reproducibility: Uses Pinned Environments, Containers, Deterministic Seeds, Explicit Configurations, Cli Workflows, Versioned Data, And Automated Verification.
Validation: Designs Objective Correctness Tests, Reference Implementations, Numerical Tolerances, Performance Thresholds, And Regression Benchmarks.
Systems Depth: Comfortable Debugging Across Python, Framework Internals, Gpu Execution, Distributed Communication, Data Pipelines, And Infrastructure.
Additional Backend Systems: Apache Pulsar, Argocd, Node.Js, Nestjs, Javascript, Flink, Iceberg, Delta Lake, Mlflow, Buildkite, Event Sourcing, Idempotency, Observability, Integration Testing.
Genai And Data Platforms: Rag, Embeddings, Semantic Retrieval, Chunking Strategies, Prompt Engineering, Structured Outputs, Function Calling, Tool Use, Grounding, Guardrails, Answer Evaluation, Feature Engineering, Time Series Ml, Data Quality, Data Modeling, Lakehouse Architecture, Object Storage, Model Monitoring, Cost-Aware Inference, Rest Apis.
Languages
Education
Leiden University
BSc · Data Science & Artificial Intelligence · Leiden, Netherlands · 9.8/10
Leiden University
MSc, Deep Learning, Machine Learning Systems, Numerical Optimization, Parallel Computing, Probabilistic Machine Learning, · Computer Science - Artificial Intelligence
Leiden University
BSc · Data Science & Artificial Intelligence
Statistics
Experience
Expertise
Qualifications
Profile
Frequently asked questions
Have questions? Find more information here.
Justas speaks the following languages: Lithuanian (Native), English (Advanced), Russian (Elementary).
Justas has at least 5 years of experience. During this time, Justas has worked in at least 8 different roles and for 5 different companies. The average length of individual experience is 1 year and 7 months. Note that Justas may not have shared all experience and actually has more experience.
Based on recent experience, Justas would be well-suited for roles such as: Performance Analysis Intern, Principal Engineer / Solo Developer, Cofounder & Lead Engineer.
Justas's most recent position is Performance Analysis Intern at Luminor Bank - CIO Office.
In recent years, Justas has worked for Luminor Bank - CIO Office, AlphaFlow, ServeAI, OpenAI, and Leiden University - LIACS.
Justas is most experienced in industries like Education, Information Technology, and Banking and Finance.
Justas is most experienced in business areas like Information Technology, Research and Development, and Quality Assurance. Justas also has some experience in Product Development and Business Intelligence.
Justas holds a Bachelor in Data Science & Artificial Intelligence from Leiden University.
Justas is immediately available full-time for suitable projects.
Similar freelancers
Discover other experts with similar qualifications and experience
Anton K.
Manuel P.
Mathias W.
Sebastian S.
Serge K.
Tungi D.
Laurin H.
Roman D.
Julien L.
Tobias J.
Experts recently working on similar projects
Freelancers with hands-on experience in comparable project as a Performance Analysis Intern
Karolis J.
Philipp G.
Germo G.
Daryoosh D.
Suyash S.
Alex O.
Mantas S.
Christiane N.
Mario T.
William N.
