Kian Heydari

TD Securities · Credit Risk Reporting · Toronto, ON

• Design, maintain, and optimize ETL pipelines in Python and Microsoft SQL Server that feed credit risk reporting.
• Migrated legacy ETL applications to Azure, replacing file-watcher triggers with semantic dependencies between jobs: an easier-to-follow flow, fewer failures, and ~50% fewer jobs.
• Automate build, test, and deployment with GitHub Actions CI/CD, and keep code secure and maintainable with Snyk, Veracode, and SonarQube.
• Volunteer on the ML Tooling team, focusing on agentic AI and helping colleagues adopt the AI tools being rolled out. My team leads the organization by a wide margin on AI-integration dashboards.

• Python
• SQL Server
• Azure
• ETL
• GitHub Actions
• Snyk
• Veracode
• SonarQube
• Agentic AI

Carleton University · Supervised by Prof. Ali Arya

• Researched LLM translation for low-resource Indigenous languages such as Northern Tutchone, which mainstream models have seen little or no data for.
• Built a training-free agentic translation pipeline and a knowledge distillation pipeline for small, locally run translation models (see Projects).
• Wrote four literature reviews: LLM capabilities and efficiency; adapting generative models to low-resource languages; retrieving in-context examples without vector indexes; and grammar induction.
• Compared mT5, mBART, XLM-R adaptations (XLM-SWCM, XLM-V), NLLB, BLOOMZ, ByT5, language-specific (PhoBERT, AfriBERTa), and specialized models (FuxiMT, DeltaLM).

• LLMs
• Low-resource NLP
• Knowledge distillation
• Agentic AI

Revolution Data Platforms

• Developed a real-time customer service call management system using Python, FastAPI, and WebSockets, with end-to-end voice data pipelines.
• Prototyped voice memo transcription with structured data extraction and automated task management based on content severity.
• Optimized multi-class text classification through prompt engineering and k-fold validation.
• Improved data quality through web scraping, transformation, and preprocessing in Python.

• Python
• FastAPI
• WebSockets
• Speech-to-text
• Prompt engineering

Paphus Solutions

• Developed chatbot APIs across multiple platforms using Python and RESTful services.
• Optimized deep learning models for custom question answering.
• Integrated Python analytics pipelines with cloud services through optimized ETL workflows.
• Migrated a Unity-based Android application to WebGL with asynchronous connection management.

• Python
• Chatbots
• Deep learning
• REST APIs
• ETL
• Cloud

Carleton University

• Taught Python workshops on data manipulation, pandas, and ETL concepts.
• Guided students through ER modeling and SQL, emphasizing data integrity and query optimization.

Carleton University

• Built a Python data generation and transformation pipeline for deep learning research.
• Designed data preprocessing workflows and custom metrics for model evaluation.

• Python
• Deep learning
• Data pipelines

Amirkabir University of Technology

• Developed a Python anomaly detection system for time-series data, with feature extraction.
• Created simulation frameworks to generate and validate synthetic training data.

• Python
• Time series
• Anomaly detection

Carleton University · Specialization in Data Science

Amirkabir University of Technology

Training-free, few-shot translation into languages the base LLM was not trained on, inspired by “Hire a Linguist!”. Parallel agents pull relevant dictionary entries, grammar rules, and examples; a translator agent and an LLM judge run a self-correction loop. Adds a map-reduce mode for large documents and a memory layer that learns from judge feedback.

• LangGraph
• DSPy
• Instructor
• LiteLLM
• Streamlit

Teacher LLMs (Mistral, Gemma 2) enrich a small English–Northern Tutchone parallel corpus with paraphrases, context, and longer sequences; multilingual students (mT5, M2M100, mBART-50) are fine-tuned fully or with LoRA in both directions, with BLEU-based early stopping, grid search, and MLflow tracking.

• PyTorch
• Hugging Face
• LoRA
• MLflow
• Ollama

Lightweight RAG pipeline in Python for domain-specific data.

• Python
• RAG
• LLMs

Search engine from scratch in Python, with custom indexing.

• Python
• pandas
• Information retrieval

pandas/NumPy pipeline reaching 93.61% classification accuracy.

• Python
• NumPy
• NLP