Artem Bakalinskii
Senior Applied ML Engineer · External Applied ML R&D
Have a large dataset and a difficult problem your team doesn't yet know how to solve?
I investigate the data, test the critical hypotheses and determine what is worth building before you commit to a full implementation.
Commercial offer
Typical scope: one research question · 5–7 business days · representative data · experiments · written findings · next-step recommendation.
Experience
Large-scale Applied ML R&D across AdTech and marketplace environments.
Senior ML Engineer — 2021–2024
One of the leading traffic verification and DSP platforms in the Russian market.
Most of my large-scale production ML experience was developed here: identity resolution, behavioral modeling, traffic verification and advertising ML — from poorly defined problems and algorithms designed from scratch to production rollout, automated pipelines and long-term monitoring.
Applied ML / R&D — 2025–2026
Russia's largest classifieds platform.
R&D for the external advertising network — expanding experience beyond a dedicated AdTech platform into a large marketplace environment with its own user, behavioral and advertising data.
Across both environments, my strongest role has been at the research-heavy end of Applied ML: taking large, messy datasets and unclear technical problems and turning them into tested algorithms and practical systems.
What I do
The most useful work often starts before model selection — understanding whether the data contains the necessary signal, what the actual constraints are, and which technical direction is worth pursuing.
How I work
The goal is not to explore everything. The goal is to collect enough evidence to make the next technical decision.
Selected work
First the proof. Details are available if you want the technical depth.
Up to ~200M users/hour · 3–4 years in production
Built from scratch a production identity resolution system for advertising infrastructure with large volumes of noisy and unstable user identifiers — no reliable global ground truth, shared IPs, bots and cascading false merges.
Fully automated hourly pipeline with persistent graph state, used for traffic verification, retargeting, audiences and downstream ML.
Identity resolution at this scale is a dynamic graph-control problem, not simply pairwise similarity.
Several billion UIDs · AUC ~0.75 → 0.83–0.85 · ~1.5–2× usable predictions
Built a stable user representation from web and in-app behavior histories (Word2Vec, 100-d embeddings, CatBoost) for gender, age-group, targeting and traffic verification — later reused by other teams.
An exploratory / Kaggle-style experiment helped shape the embedding approach that was then adapted for production at billions of UIDs.
Strong behavioral representation mattered more than adding feature complexity.
Behavior history → Word2Vec domain/app embeddings (100-d) → user embedding → CatBoost → class-specific thresholds
Beyond production
I use Kaggle competitions and exploratory projects to enter unfamiliar ML domains, test approaches quickly and transfer the strongest ideas into production.
Experiments have included behavioral modeling, 3D medical imaging, MRI and ECG data, segmentation and other applied ML problems.
One Kaggle-style experiment directly influenced the behavioral embedding approach later adapted for a production system on billions of user IDs.
The common pattern is the same: understand an unfamiliar data problem, test hypotheses quickly and find a practical technical direction.
Engagement details
Before committing to a full ML implementation, determine what your data can actually support.
A focused R&D engagement that answers one primary technical question — not a full product build.
5–7 business days
Sensitive data can remain inside the client's infrastructure. I can work with anonymized samples or within the client's environment.
Enough evidence to proceed with development.
Signal exists, but more research is needed.
Original framing is weak; a better path was found.
Data or economics do not justify further development.
A well-supported negative result is also a successful R&D outcome.
For suitable problems, I can start with a limited preliminary review of the problem and available data before defining the full research scope.
About
I'm a Senior Applied ML Engineer focused on research-heavy, data-intensive problems.
My strongest production background is in AdTech, identity resolution and behavioral modeling, but the underlying expertise is broader: investigating large, noisy datasets and finding practical ML approaches where the correct solution is not obvious upfront.
Contact
If you have data but are not yet sure what can realistically be built from it, send me a short description of the problem.