Machine Learning
Transformer models for sequences, gradient boosting (CatBoost, LightGBM), time series and forecasting, explainable AI (attention maps), leak-free validation.
Open to an internship / junior role
Hi, I'm NuIIX
ML Engineer · AI Developer · Data Scientist
I'm Maxim, a fourth-year student at SibSUTIS working on machine learning in Python: from data preparation and model training to packaging models into a working service.
- EducationSibSUTIS · class of 2027
- CertificateSamsung Innovation Campus
- LocationNovosibirsk · UTC+7
01 / About
From raw data to a working service
Training a model is only half the job. I care about bringing it to a state where people can actually use it.
I work on machine learning and Python development: preparing data, training models, validating them without leakage and wrapping them into FastAPI services.
I'm looking for an internship or a junior position as an ML Engineer, AI Developer or Data Scientist. Meanwhile I'm working on my thesis and a mobile app in Flutter.
- Education
- SibSUTIS · 09.03.01 Informatics and Computer Engineering · class of 2027
- Certificate
- Samsung Innovation Campus — Artificial Intelligence and ML (with distinction, 2026)
- Focus
- ML projects, thesis, a Flutter mobile app
- Growing
- ML/DL, backend, Kubernetes
- Location
- Novosibirsk · UTC+7
Backend & MLOps
FastAPI services and REST APIs, Docker, Nginx, CI/CD, PostgreSQL, Redis, Prometheus monitoring, API security (IDOR / BOLA).
02 / Projects
Selected work
Two key projects, from the problem and approach to honest results and limitations.
NeuralIDOR-Det
Deep Learning · SecurityDetecting anomalous behaviour of users and bots in API traffic, primarily IDOR-like requests to resources owned by someone else.
Approach
A Transformer Encoder analyses sequences of API calls while an MLP handles metadata (IP, user_id, request count). Features are combined through late fusion.
An autoencoder looks for unknown attack types without labels.
Explainability
Attention maps show which API calls influenced the model's decision, so the result can be inspected rather than taken on faith.
- 0.784Accuracy
- 0.870F1
- 0.666ROC-AUC
Honest about limitations. The CatBoost baseline reaches ROC-AUC 0.893, so the hybrid model does not beat it on this metric yet. Causes and the improvement plan are described in the project README.
Semantic Gardener
Obsidian plugin · NLPAn Obsidian plugin that finds semantically duplicated notes and carefully merges them into atomic concepts.
Approach
Local embeddings (multilingual MiniLM, ONNX) are computed in a Web Worker. Clustering uses cosine similarity and each duplicate is double-checked with Gemini (BYOK).
UX & privacy
Human in the loop: word-level diff, accept / edit / reject, one-click undo of the last refactoring.
Vault text never leaves the machine during scanning.
Other public repositories
- py-cointracker CLI crypto tracker: collects data from coinmarketcap.com, stores it and exports to CSV / JSON / XML. Python
- Yeahub-Obsidian-Parser Parses questions and collections from the Yeahub API and formats them for Obsidian.md. Python
- lithology-classifier Classifies lithological rock types from a set of well-log curves. Jupyter
- RedisTutorialProject Redis as a cache in ASP.NET Core MVC with database and RESTful service examples. C#
- SimpleExampleAuth A minimal authentication and authorization example. C#
- ASCII_Art_Generator Converts an image into ASCII art. C#
Some projects are private: ML forecasting, FastAPI services, Flutter apps. Details on request.
03 / Journey
From web services to transformers
-
2024
C# and ASP.NET MVC
Redis caching, authorization, console utilities.
-
2025
Python and data engineering
Parsers, crypto-asset tracking, working with APIs.
-
2026
ML and Deep Learning
Samsung Innovation Campus, anomaly detection in API traffic, Flutter and an Obsidian plugin.
04 / Stack
Technologies
Languages
ML / Data Science
Backend & Web
Databases & streaming
DevOps
Tools
05 / Contact
Let's talk
I'm looking for an ML / AI internship or a junior role. Telegram is the fastest way to reach me. If you'd like to see the code, it's all on GitHub.