About

Machine Learning

Deep learning, reinforcement learning, multi-agent systems, causal inference, computer vision, NLP, search and ranking, conversational AI.

Learning for Physical Systems

RL policies for embodied AI, sim-to-real transfer, vision-sensory-motor policies for dexterous manipulation, policy robustness. NVIDIA Isaac Lab, Gazebo, ROS.

Scientific ML

Surrogate modeling of physical simulations, climate and geophysical time series, forecasting under sparse and irregular observation.

Experimentation

A/B testing infrastructure, causal-inference systems, evaluation design for research-to-production transitions.

Tools & Systems

Python, C++, PyTorch, TensorFlow, SQL, AWS, GCP, Elasticsearch, Airflow, Docker, Slurm, dbt, big-data pipelines.

• Intrinsic AI for Industry Challenge (Alphabet): developed vision-sensory-motor policies for dexterous manipulation on server-tray wiring assembly, in Gazebo and NVIDIA Isaac Lab. Placed 56th globally among 160+ specialized teams.
• Confidential robotics research (stealth): architecting high-precision reinforcement-learning policies for embodied AI, leading the transition from simulation to hardware with a focus on policy robustness.
• Humanoid robotics education (Unitree G1), volunteer: helped establish a humanoid-robotics program for high-school students; part of the team that built a pipeline retargeting recorded human motion onto the G1, and supervised students through deployment on real hardware.

• Architected and deployed a customer analytics engine for a banking client that automatically identifies behavioral patterns, loss reasons, and re-engagement opportunities across the customer lifecycle — surfacing hundreds of thousands of dollars in recoverable revenue.
• Built a data-grounded insight-generation layer for C-suite stakeholders, with every generated claim traceable to its supporting evidence.
• Retrieval-augmented generation over vector databases, deployed on GCP with a multi-tenant architecture for privacy compliance.

• Founded an authorized AI likeness platform creating monetizable, personalized AI versions of creators; designed the core ML architecture end-to-end.
• Designed responsive personality models that learn individual communication style, tone, and behavioral patterns from creator data.
• Owned every technical decision across ML, infrastructure, and product as sole technical founder.

• Designed a prediction-aggregation system for a computer-vision pipeline detecting plant disease, improving detection robustness and reducing manual intervention.
• Guided the transition of research prototypes into production-ready systems through architecture review and design.

• Built personal projects end-to-end: two games, a social media platform, and a dating platform.

• Promoted from ML Scientist to Senior ML Scientist.
• Autocomplete: owned the deliverable across product, frontend, backend and QA on a compressed timeline. Personally implemented the ML experimentation and inference plus the backend data-processing and serving code — retrieval, relevance tuning and re-ranking on Elasticsearch, serving millions of shoppers monthly. A/B test showed a 30% lift in mobile search engagement.
• Led a team of 4 on causal inference analysis, and designed the causal-inference system measuring the business impact of algorithmic interventions. The statistical framework was adopted cross-functionally for customer-experience decisions.
• Architected core components of the "Data Vault", a centralized platform ingesting 6 TB/day as the single source of truth for personalization, experimentation, and analytics.

• FishNet: built an end-to-end deep-learning system processing hundreds of thousands of underwater images to classify 150 fish species and estimate size, replacing hundreds of hours of manual annotation. Published, IEEE.
• First-authored a deep-learning surrogate for Atlantic sea-surface temperature forecasting: a CNN with U-Net/ResNet-style skip connections that matched state-of-the-art dynamical forecast models at a fraction of their cost, and beat persistence baselines at 1–6 month lead times. Architecture selected by Bayesian optimization over 400 trained networks, run multi-GPU on Slurm. Published, AAAI.
• Built a neural network for galactic cosmic ray intensity estimation, approximating classical differential-equation solvers roughly 1,000,000× faster at under 3% relative error, while removing their numerical instabilities.
• Contributed quantitative analysis to a marine spatial-management study published in Nature Communications.

• Built an end-to-end data science pipeline — scraping, storage, preprocessing, analytics — for housing-market forecasting across 300,000 properties, producing daily forecasts.

• Taught R for Data Analysis and Visualization; supervised 11 student projects.

• Teaching assistant for Statistics, Microeconomics and Game Theory alongside the M.A. Economics: ran office hours and graded for classes of up to 150 students.

• Prototyped an NLP system for automated service-request handling across 5,000+ categories for the largest bank in Eastern Europe.
• Built anomaly detection for request-volume spikes.

• Tutored Mathematics and Physics from middle-school through college level, including exam and olympiad preparation.

Mots'oehli, M., Nikolaev, A., et al. — IEEE, 2024

Nikolaev, A., Richter, I., & Sadowski, P. — AAAI Workshop (CEUR-WS Vol. 2587), 2020

Co-author — Nature Communications, 2020