Open to AI/ML & Software Engineering internships

Alibek Berik

AI & Machine Learning Engineer

I build end-to-end ML systems — from raw data and feature engineering to models deployed behind real APIs and live dashboards.

See my projects → Download CV

alibek.py

from career import MLEngineer

alibek = MLEngineer(
    focus=["LLMs & RAG", "Computer Vision", "NLP"],
    stack=["PyTorch", "ChromaDB", "FastAPI", "Docker"],
    speaks=["English", "Русский", "Қазақша"],
    status="open_to_internships",
)

alibek.ship()  # data → model → production

1,000+ real user requests through my production AI pipelines
800K+ transactions engineered into churn features
40→67% churn recall after fixing class imbalance
97% CNN test accuracy on 10,000 images

About me

Turning data into working products

I’m a second-year Artificial Intelligence student at Xiamen University Malaysia who learns by building. As an AI/ML Engineer at GlowUp Face, I built inference pipelines combining open-source computer vision models with the OpenAI API, processing 1,000+ real user requests for facial, skin and hairstyle analysis.

Outside work I’ve shipped projects across LLMs, NLP, computer vision and tabular ML, and I take every one past the notebook: a Dockerised RAG system that cites its sources and refuses to hallucinate, SQL feature pipelines over 800,000+ real transactions, a fine-tuned DistilBERT served through FastAPI, and live dashboards on Streamlit Cloud. I debug models until they actually generalise, not just fit the training data, and I document projects so results are reproducible.

I’m looking for an AI/ML or Software Engineering internship where I can contribute to real intelligent systems and grow alongside experienced engineers.

Why hire me

What I bring to your team

I've already shipped AI to real users

At GlowUp Face I built computer-vision and OpenAI inference pipelines that handled 1,000+ real user requests — and every personal project I build ends up deployed, from Streamlit Cloud dashboards to FastAPI services.

I take data seriously

I caught and fixed target leakage in my churn pipeline, and I handle class imbalance deliberately with SMOTE — measuring the precision tradeoff instead of hiding it.

I improve things measurably

Churn recall from 40% to 67%. House-price R² from 0.35 to 0.60. I iterate with metrics and experiments, not guesses.

I learn fast and communicate clearly

In about a year I've gone from regression to CNNs, transformers, RAG systems and tool-using agents, mostly self-taught. I work in English, Russian and Kazakh and write clear, honest READMEs.

Skills

My toolkit

LLMs & Agents

RAG
ChromaDB
sentence-transformers
OpenAI API
Groq API
smolagents
Prompt engineering

Deep Learning

PyTorch
TensorFlow / Keras
Hugging Face Transformers
CNNs
Fine-tuning

Machine Learning

Scikit-learn
XGBoost
LightGBM
imbalanced-learn (SMOTE)
Naive Bayes
Hyperparameter tuning

Deployment

FastAPI
Docker
Streamlit Cloud
Hugging Face Hub
Inference pipelines

Data

Python
SQL
Pandas
NumPy
SQLite
Feature engineering

Tools & soft skills

Git
Linux
Jupyter
Google Colab
Trilingual communication
Technical writing

Experience

Where I've worked

AI/ML Engineer

GlowUp Face

Aug 2026 – Oct 2026
- Engineered and optimised AI inference pipelines combining open-source computer vision models with the OpenAI API, processing 1,000+ user facial, skin and hairstyle analysis requests.
- Streamlined backend workflows and prompt-engineering pipelines, reducing image-analysis latency and improving response accuracy for a platform scaling to thousands of active users.
- Collaborated on product architecture and feature deployment for an AI-driven consumer platform, building core machine-learning capabilities from scratch.

Projects

Things I've built and shipped

LLMs · RAG · Vector Search · API 2026

Kazakh Culture RAG Q&A System

A retrieval-augmented Q&A system that answers questions about Kazakh culture from real Wikipedia sources — citing them, and refusing to answer when the knowledge base doesn't cover the question.

6-stage pipeline · fetch → chunk → embed → store → retrieve → generate

ChromaDB
sentence-transformers
Groq LLM API
FastAPI
Docker

How I built it

Built a full RAG pipeline: fetches Wikipedia articles (nomadic pastoralism, yurts, dombra, Nauryz, cuisine, eagle hunting, the Kazakh Khanate) and splits them into overlapping ~500-word chunks.
Embedded chunks with sentence-transformers (all-MiniLM-L6-v2) and stored them in a ChromaDB vector database for semantic retrieval.
Generated answers with an LLM via the Groq API, grounded in retrieved context and citing sources — out-of-scope questions get “I don’t have enough information” instead of a hallucination.
Served through a FastAPI endpoint and containerised with Docker; the index rebuilds from scratch in under a minute.

GitHub ↗

SQL · ML Pipeline · Deployment 2026

Customer Churn Prediction System

Flags e-commerce customers who are about to churn, using 800K+ real transactions from the UCI Online Retail II dataset, served as a live risk-lookup dashboard.

40% → 67% churn recall · AUC-ROC 0.759

SQL
XGBoost
SMOTE
Pandas
Streamlit

How I built it

Engineered RFM (Recency, Frequency, Monetary) features with SQL CTEs, joins and aggregations over 800,000+ real e-commerce transactions (Dec 2009 – Dec 2011).
Identified and resolved a target leakage issue in the initial feature set before it could inflate results.
Trained an XGBoost classifier and compared a baseline against SMOTE balancing: recall on churned customers rose from 0.40 to 0.67 (+68%) while precision moved from 0.52 to 0.47 — a deliberate, measured tradeoff.
Deployed as an interactive Streamlit dashboard for real-time customer churn-risk lookup.

Live demo ↗ GitHub ↗

NLP · Fine-tuned Transformer · REST API 2026

Sentiment Classifier

A fine-tuned DistilBERT model that classifies movie-review sentiment, served in real time through a FastAPI backend and published on Hugging Face Hub.

0.866 F1 score · 85.9% accuracy

PyTorch
Hugging Face Transformers
DistilBERT
FastAPI

How I built it

Fine-tuned distilbert-base-uncased on IMDb (3,000 training / 1,000 test reviews) on a Colab T4 GPU, reaching 85.9% accuracy and a 0.866 F1 score.
Built a FastAPI backend serving real-time predictions, with an interactive Swagger UI for testing.
Published the trained model to Hugging Face Hub and documented free-tier hosting constraints transparently in the README.

Model on Hugging Face ↗ GitHub ↗

Regression · Feature Engineering 2025

Malaysia House Price Prediction

A full regression pipeline predicting Malaysian property prices from a 5,000+ record real-estate dataset.

0.35 → 0.60 R² score (+71%)

LightGBM
Scikit-learn
Pandas
NumPy

How I built it

Engineered custom features — estimated total price and location clusters — and applied LightGBM gradient boosting, improving R² from 0.35 to 0.60 (a 71% relative improvement).
Ran the full pipeline: data cleaning, outlier removal, label encoding, cross-validation and hyperparameter tuning.

GitHub ↗

LLM Agents · Tool Calling 2026

Tool-Using AI Agent

An AI agent that reasons step by step and calls real tools — live web search and timezone lookups — so its answers are grounded in actual data instead of guesses.

smolagents
Hugging Face Inference API
Ollama
Python

How I built it

Built with Hugging Face smolagents using a Thought → Action → Observation reasoning loop.
Gave the model real tools: a DuckDuckGo web search and a custom get_current_time_in_timezone function, so questions like “What time is it in Tokyo right now?” are answered from live data.
Ran models through both the Hugging Face Inference API and local Ollama, and documented why next-token prediction alone leads to hallucination.

GitHub ↗

Deep Learning · Computer Vision 2025

MNIST Digit Classifier

A convolutional neural network that recognises handwritten digits, with a visual analysis of prediction confidence.

97% test accuracy

TensorFlow
Keras
CNN
NumPy

How I built it

Architected and trained a CNN in TensorFlow/Keras on 60,000 samples, reaching 97% test accuracy on 10,000 held-out images.
Implemented image normalisation, dropout regularisation and a visualisation pipeline for prediction-confidence analysis.

GitHub ↗

NLP · Text Classification 2025

Spam vs Ham Classifier

A message classifier that separates spam from legitimate messages, with an interactive terminal interface for real-time predictions.

~95% accuracy on 10,000+ messages

Scikit-learn
Naive Bayes
Bag-of-Words
Pandas

How I built it

Extracted text features with a Bag-of-Words model (CountVectorizer) and trained a Multinomial Naive Bayes classifier on 10,000+ labelled messages.
Evaluated with accuracy, a confusion matrix and a full classification report.
Added an interactive terminal interface for classifying new messages in real time.

GitHub ↗