Kourosh

Data & AI Specialist

Based in Milan, Italy

M.Sc. student · Politecnico di Milano

Scroll to explore

I work across data and AI, turning data into product insights, building data pipelines, and developing machine learning systems.

Currently pursuing a Master of Science in Computer Science and Engineering (Artificial Intelligence) at Politecnico di Milano.

Education

  1. Milan, Italy

    Master of Science (M.Sc.) in Computer Science and Engineering (Artificial Intelligence)

    Course project · Natural Language Processing ↗

    Built an end-to-end question-answering system using RAG-Instruct. Compared a Transformer from scratch, fine-tuned Flan-T5 and LLaMA-3.2-1B.

    • Fine-tuned LLaMA-3.2-1B with QLoRA and 4-bit quantization
    • Evaluated with BLEU, ROUGE and BERTScore; added Whisper speech input, text-to-speech and a Gradio feedback interface

    Artificial
    Intelligence    Current M.Sc. studies

  2. Tehran, Iran

    Bachelor of Science (B.Sc.) in Computer Engineering (Computer Architecture)

    Thesis

    Design and Implementation of an Arduino-Based Electronic Throttle Valve ↗

    • Integrated Arduino-based control logic with an electronic throttle mechanism
    • Built a hardware-software prototype as a foundation for future autonomous vehicle and robotics experiments

    Computer
    Architecture    Computer Engineering

Career

  1. Feb 2023 — Dec 2024

    Built the data infrastructure behind analytics, business intelligence and recommendation systems for Shopify e-commerce.

    Python / SQLBigQuery · Power BI · Metabase · Django

  2. Jan 2019 — Aug 2022

    Part-time, hybrid coaching in a behavioral weight-management setting.

    • Coached 50+ participants through four-month engagements and held individual sessions with 300+ people.
    • Used active listening and empathetic communication to support reflection on values and commitments; coordinated activities with fellow coaches.

    50+participants in four-month coaching engagements

  3. Oct 2019 — Aug 2020

    Research at the Robust and Interpretable Machine Learning Lab, focused on noisy MEG neuroimaging data. Research project ↗

    • Researched adversarial robustness and interpretability for CNN-based classification models.
    • Developed PyTorch training and evaluation pipelines to compare clean and adversarially perturbed conditions; used Grad-CAM to study model attention.

    PyTorchCNNs · adversarial robustness · interpretability

Selected Works

Professional work, academic research and personal projects.

Fast Bundle · Data infrastructure & analytics

Data Infrastructure & BI Platform

Built Fast Bundle’s data infrastructure end to end, from ETL/ELT pipelines to automated reporting and business dashboards.

  • Python
  • SQL
  • BigQuery
  • Power BI
  • Metabase
Project details
  • Designed, built and maintained data infrastructure and ETL/ELT pipelines with Python, SQL and BigQuery.
  • Automated data ingestion, transformation and reporting workflows; centralized transformation logic and reduced manual reporting steps.
  • Created interactive dashboards and automated BI reports in Power BI and Metabase for product, marketing and leadership teams.
  • Built KPI tracking systems and worked with stakeholders on metric definitions and OKR alignment.

Fast Bundle · Professional project

Product Recommendation Microservice

A Django-based recommender connecting machine learning models with Shopify product workflows.

  • Django
  • Python
  • Machine learning
  • External APIs
Project details
  • Developed, deployed and maintained a recommender microservice for personalized product recommendations.
  • Connected model logic with external APIs and product workflows.
  • Built supporting data pipelines for model inputs and model behavior evaluation.

Sharif University of Technology · Research

Robust & Interpretable Models for MEG

Research into adversarial robustness and interpretability for CNN-based neuroimaging classification.

  • Python
  • PyTorch
  • CNNs
  • Grad-CAM
Project details
  • Researched neural models applied to noisy MEG neuroimaging classification tasks.
  • Built training and evaluation pipelines comparing clean and adversarially perturbed conditions; used interpretability methods to study model attention.

Sharif University of Technology · B.Sc. thesis

Arduino-Based Electronic Throttle Valve

An embedded hardware-software prototype integrating Arduino control with an electronic throttle mechanism.

  • Arduino
  • Embedded systems
  • Control prototyping
Project details
  • Designed and implemented the electronic throttle valve as a bachelor’s thesis.
  • Integrated Arduino-based control logic with automotive hardware.
  • Built a foundation for future autonomous vehicle and robotics experiments.

Neuromatch Academy · Team project · 2023

Multi-Agent Reinforcement Learning

A Neuromatch Academy team project exploring multi-agent reinforcement learning and model interpretability.

  • Reinforcement learning
  • Interpretability
Project details
  • Completed as an academic team project at Neuromatch Academy in 2023, within deep learning and machine learning.

Politecnico di Milano · NLP course project

Question Answering with LLM Fine-Tuning

An end-to-end QA project comparing a Transformer from scratch, Flan-T5 and LLaMA-3.2-1B.

  • Transformers
  • QLoRA
  • PEFT / LoRA
  • Whisper
  • Gradio
Project details
  • Compared a Transformer built from scratch, fine-tuned Flan-T5 and LLaMA-3.2-1B with zero-shot and one-shot prompting.
  • Fine-tuned LLaMA-3.2-1B using QLoRA, 4-bit bitsandbytes quantization and Hugging Face Transformers.
  • Evaluated with BLEU, ROUGE and BERTScore; added speech input/output and a Gradio user-feedback interface.

Personal project · Desktop AI

Telegram Chat Summarizer

A Python desktop app for read and unread chat summaries, sentiment and per-user interaction analysis.

  • Python
  • Telethon
  • customtkinter
  • OpenAI API
  • Ollama / Qwen3
Project details
  • Summarizes one conversation at a time, with unread-chat browsing and separate read/unread summaries.
  • Switches between cloud inference and a local Qwen3 8B fallback, with token counting, retries and structured error handling.
  • Supports multilingual prompts, including Persian, and Telegram phone, code and 2FA login within the app.

Personal project · Developer tooling

Obsidian Password-Protect Plugin

A working TypeScript plugin for password-protecting and encrypting individual Obsidian notes.

  • TypeScript
  • Obsidian API
  • CryptoJS
  • esbuild
Project details
  • Implemented locking, unlocking, temporary unlock and permanent removal of note protection.
  • Used CryptoJS AES encryption and SHA-256 password hashing, with esbuild-based packaging.
  • Delivered the plugin end to end with AI coding agents while learning TypeScript.

Contact · Milan, Italy

Let’s work
together.

kourosh.zavvari@gmail.com