Software engineer in Astana, Kazakhstan. MSc student at Astana IT University, researching machine learning for atmospheric monitoring. Coding mentor at Tomorrow School, where I design and teach the ML/AI curriculum.
Most of my day-to-day work is backend engineering in Go and Python; my research work is applied ML with an emphasis on explainability and on being honest about what a model cannot do.
Heavy-metal air pollution forecasting with explainable ML — ongoing MSc work on 15 years of state monitoring records from Almaty (2006–2020). A reproducible pipeline reconstructs a modelling table from 32 heterogeneous source workbooks, benchmarks 10 tuned regressors, and attributes predictions with SHAP and LIME. Extensions beyond the reference methodology: LSTM/GRU comparison against naive baselines, quantile regression with calibration checks, and a Chow test for structural breaks in the series.
Two of the findings are negative, and the write-up reports them as such: recurrent networks failed to beat a window-mean baseline on this sparsely-sampled, non-stationary series, and the quantile intervals under-cover by 8–12 percentage points.
Crude oil classification from laboratory assays — a smaller study on the open NOAA Oil Library, and mostly a lesson in checking your own results. The first version scored 98.4%; auditing it showed the label was thresholded API gravity while density sat in the feature matrix, and API is a closed-form function of density — a two-line formula with no model at all reproduces the label on 99.35% of rows. The corpus also reported every property in several units without normalisation. With the leaking columns removed and units fixed, the honest figure is 85.5% against a 53.2% baseline.
AI code review assistant — a FastAPI service that reviews GitLab merge requests with an LLM and posts findings as inline comments anchored to the diff. The engineering is mostly in the failure modes: constant-time webhook authentication, guards against the bot retriggering itself, bounded-concurrency chunking of large diffs, provider-enforced JSON with retry and backoff, deduplication of findings repeated across chunks, and a summary that states how many chunks were actually reviewed rather than hiding partial coverage.
Intervu — an adaptive technical-interview platform I am designing: LLM-based extraction of claims from interview transcripts, a Bayesian competence model with time decay, and a strict separation between prediction and decision so that no score is produced during the interview itself. Currently in the design phase — 17 ADRs, an OpenAPI contract, and a PostgreSQL/pgvector schema whose invariants are enforced in the database rather than by convention.
ML/AI curriculum design — author of a seven-project production-ML branch plus six electives, taught at Tomorrow School: MLOps serving platform (FastAPI + ONNX Runtime, Prometheus/Grafana, a versioned model registry), recommender systems, LLM engineering with retrieval and agents, document AI, vision tracking, and predictive analytics with SHAP/Fairlearn interpretability. Every project ships a runnable service and a report stating where the system is weak.
| Project | What it is |
|---|---|
| almaty-heavy-metals-ml | ML pipeline for air-quality forecasting: 10-model benchmark, SHAP/LIME, LSTM/GRU, quantile regression, structural-break testing |
| ai-code-review-assistant | LLM reviewer for GitLab merge requests. Async FastAPI, diff chunking, JSON-mode output with retries, deduplicated inline comments |
| crude-oil-classification | Oil grading on open NOAA assays: target-leakage audit, unit normalisation, honest baseline, Streamlit app |
| clipflow | Native macOS clipboard manager. Swift 6 with strict concurrency, SwiftUI + AppKit, SwiftData. Fully local |
| forum | Go web application with sessions, posts, comments and filtering, built on the standard library |
| Multi-File-Mailer | Go API for bulk multipart file delivery and ZIP archiving |
| tcp-chat | Concurrent TCP chat server in Go: multiple clients, broadcast and private messaging |
Not public here: an internal Go + React dashboard I wrote and maintain — multi-campus RBAC, backend-enforced time tracking with network allowlisting and heartbeat limits, secure OAuth2 sessions, and an immutable audit log. I also contribute to an Ubuntu fleet-provisioning system for lab machines built on overlayroot and Ansible.
Languages — Go, Python, Swift, JavaScript, SQL, Bash
ML / Data — scikit-learn, XGBoost, CatBoost, TensorFlow/Keras, SHAP, LIME, pandas, NumPy, SciPy
Backend — FastAPI, PostgreSQL, SQLite, Redis, chi, REST, GraphQL, TCP/WebSocket, OAuth2
Other — Docker, Ansible, React, SwiftUI, Git
Explainable ML for regulatory and safety-critical use, uncertainty quantification, LLM systems where non-deterministic components are isolated behind reproducible ones, and teaching engineers to build ML systems that survive contact with production.
Reach me at bbvbatvv@gmail.com.