I train models
and take them
to production

Problem

Data on athletes, coaches, venues and results was stored across dozens of separate files and systems. Preparing a single analytics report took weeks of manual work. Leadership had no holistic view of the industry.

Solution

• Designed a relational DB (15+ tables, PostgreSQL) to unify all sources
• Geocoding — placed 3,600+ pools across Russia on an interactive map
• A composite index of regions from 124 absolute and 60 aggregated metrics
• Dashboard with regional rankings, heatmaps and filters
• Index system: infrastructure availability, staffing, participation

Result

Leadership got a decision-making tool: where infrastructure is overloaded, where it's underused, which regions are drivers. The ability to model the effect of investments.

Problem

The inter-ministerial program "Swimming for All" needed an official public portal — content, maps, data and applications — on a hard deadline, not in a quarter.

Solution

• A video-course module: 2 courses, 20 lessons, watchable right on the portal
• A pool map with "Near me" geo-search
• A dashboard of program indicators broken down by region
• An application form for joining the program, with automatic notification of the officials in charge
• A full admin panel: content, news and data are managed without a developer

Result

From first commit to a live public portal in 7 calendar days — in Russian and English, with content and data maintained through the admin panel ever since.

Problem

There was no single, comparable picture of how countries develop mass swimming education and water safety. The data was scattered across national federations, ministries and methodology documents in dozens of languages.

Solution

• Designed the SDI index — 8 weighted criteria (0–5 scale, weights sum to 100) with a transparent formula
• Deep-research profiles for 198 countries from official sources: federations, ministries, statistics, methodologies
• Interactive D3 + TopoJSON world map: country search, tooltips, side panel and colour legend
• A drill-down page per country plus an open, documented scoring methodology

Result

A single tool to compare countries, separate leaders from laggards and ground decisions in transparent, reproducible scoring.

Problem

The federation funds the training of hundreds of athletes, but the budget is limited. Coaches pick candidates intuitively — subjectively and without common criteria.

Solution

• Compared four ML algorithms and chose the best by cross-validation (LightGBM)
• Tuned hyperparameters over 40 iterations (RandomizedSearchCV)
• Used SHAP to surface and rank the handful of signals that actually move the outcome
• Checked class imbalance effect (SMOTE/SMOTENC) — the ceiling is set by data quality

Result

The model flags future medalists with 91.6% accuracy. Behind every verdict is a multi-factor read of the athlete's trajectory, and each prediction ships with its own explanation — while the decisive signals stay under the hood.

Problem

A coach judges technique by eye — but misses micro-asymmetries that reduce efficiency and raise injury risk. Objective numbers are needed: angles, ranges, side balance.

Solution

• 17 key body points in every frame (YOLOv8 / MediaPipe)
• Angles for each joint, broken down by movement phase
• Automatic left/right comparison — flags asymmetries
• Export to a report with charts to track progress between sessions

Result

Coaches got objective metrics instead of subjective judgement. Athletes correct movements faster and lower injury risk thanks to early imbalance diagnosis.

Problem

Lab technicians count blood cells by hand under a microscope — slow, tiring and error-prone. One image takes minutes, and there are dozens a day.

Solution

• Faster R-CNN (ResNet50-FPN v2), fine-tuned on blood-smear images
• Three classes in one pass: red cells, white cells, platelets
• Transfer learning — adapted in 10 epochs from ImageNet weights
• Each cell boxed with class and confidence

Result

The model confidently detects cells at 0.9+ confidence. One image — in a second instead of minutes of manual counting. A ready base for integration into lab systems.

Problem

Official grant reports demand a DOCX with a line-by-line financial audit, assembled from 1000+ messy scans, contracts, payment orders and estimates. By hand it's days of tedious, error-prone work.

Solution

• An orchestrator runs the process; Python scripts do the deterministic part — indexing, extraction, DOCX assembly, financial audit
• 6 parallel OCR sub-agents read scans, contracts and payment orders; 3 more parse the finances and write the narrative sections
• Confidential financial and personal data is processed by local, on-device models — nothing sensitive ever leaves the machine
• A document-linking graph and triangulated audit reconcile every figure against limits and primary documents
• Per-row Word comments, an auto "remarks to check" doc and a clean fund-ready copy — then a mandatory self-review pass

Result

1000+ source documents (~1.5 GB) → an audited, fund-ready report in a single run, every line cross-checked and commented. Days of manual work collapse into minutes.

Problem

Athletes' lab results — special-category personal data — were scattered across PDFs from different labs. Doctors needed trends over time in one place, but cloud services are a non-starter for sensitive medical data.

Solution

• Parses lab-report PDFs / XLSX from many labs into structured markers (athlete → report → test)
• Trends over time per marker, normalized against a reference-value catalog; dashboards and PDF/XLSX export
• Role-based access (admin / medic / viewer) with per-record scoping, brute-force protection and a full audit log of every action
• Closed loop by design: runs on-prem / behind a corporate VPN, isolated from the public internet; encrypted off-box backups
• A companion smartphone app for access on the go

Result

Doctors get longitudinal trends and instant reports; the organization keeps full control — medical data stays inside the perimeter, every access is logged, and nothing goes to the cloud.