AI Engineer · LLMs, apps & NLP
I build LLM tools, apps and NLP pipelines with Claude Code, Codex and Cursor, and I test what the AI writes before it ships.
Open to AI internships · Baku · Graduating Dec 2026
Aug 2026 – Present
Fine-tuned a transformer model that finds and masks personal data in documents, and built the LLM pipeline that creates and checks its training data. Now building validation pipelines for a new decision model with Claude Code.
Work under NDA. Details stay general.
Jun 2026 – Present
AI podcast-summary app, live on the App Store with 150+ downloads and 1st place at University Demo Day. I built the AI pipeline (Gemini, Whisper, Deepgram) on a Supabase backend.
May 2026 – Aug 2026
Benchmarked EdTech platforms on topics, pricing and enrollment trends, and turned the findings into positioning recommendations.
Muaataz is one of the best researchers I've ever met. He's fantastic to work with, very professional and organized. He puts together spectacular presentations and goes above and beyond in so many ways.
Sep 2025 – Feb 2026
Led the tech for a car-parts marketplace, from a Telegram bot prototype to a full marketplace website.
Accepted and presented · Best Presentation Award · to appear in ACM ICPS
Project lead · co-authored with A. Aliyev, B. Hasanov, I. Alishov, A. Huseynli, S. Rustamov
Project lead · Senior Design Project, ADA University · 2 faculty supervisors
A video deepfake detector that looks for inconsistencies across frames instead of single-image artifacts, extended across video and text modalities.
ADA University · ML capstone · team project
Predicting diabetes from 8 health readings per person. Six models, 5-fold grid search with F1 as the target, and every run tracked in MLflow.
Walkthrough on request
Led a team that built a prompt-risk guardrail prototype in one day at Cursor Hackathon Baku, then pitched it and delivered the final demo.
Six-way news topic classifier for Azerbaijani text using TF-IDF features. A linear SVM beat Naive Bayes (77.6%) and a decision tree (72.3%).
About 500k rows and 54 features. Compared classifiers, then soft voting with weak voters removed. A second experiment majority-voted 1,200 tiny trees.
Sigmoid, log loss and batch gradient descent written by hand, with a decision-boundary plot, then checked against scikit-learn on the same data.
Car prices from messy real marketplace listings. Hand-written gradient descent, scikit-learn and the normal equation, cross-checked against each other.
Grid robot with dirt, walls and movement slip. Solved with value and policy iteration, then compared against a tabular Q-learning agent.
Team computer-graphics project. A browser 3D game with custom shaders, seeded world generation, collisions and camera controls.
Notebook analysis of international results (1872–2017) covering team dominance, penalties, friendlies and top scorers.
Java Swing desktop app with authentication, CRUD, sorting and filtering, built with shared Git history and code reviews.