Amir Aghdam

• Spearheading the early development of a user-focused chatbot using OpenAI on Azure, llama,and langchain, enhancing user engagement on our real estate platform, Hominext. Reducing the Response Time by 80% and Response Time Variance by 90%
• Contributed to a Deutsche Telekom project, developing a voice assistant for scheduling doctor's appointments, utilizing Rasa and Python to improve data processes and recognition of names and diseases. We were able to land our first customer during my time on the team.
• Participated in Deutsche Telekom's Magenta project, applying machine learning techniques to classify user intents from a broad range of classes and subclasses, enhancing the accuracy of intent classification. We were able to reduce our Cloud Computation Cost by 50% by optimizing our system

• Technologies: Python, Pandas, ChatGPT, langchain, Gradio, Rasa, Docker, Jupyter, Dash
• Theory: LLMs, Chatbot, Entity Recognition, Intent Classification

• Designed & Developed an online Dashboard building tool on Azure using Python, CSS and Plotly
• Dash to facilitate creating and launching customized dashboards for our customer's (Lanxess AG)data science team
• Migrated more than 20 complex dashboards from Palantir to Microsoft PowerBI

Technologies: Python, Dash, CSS, PowerBI, Azure (Data Lake, DevOps, Blob Storage), Luigi

Theory: Data Pipeline, Dashboard Visualization, Cloud Integration.

Sotoon, a new branch of Hezardastan Group, focused on Infrastructure & AI as a service

• Investigated SoTA text processing solutions in the field of Natural Language Processing, paving the way for implementation in Sotoon's future text processing projects and holding biweekly seminars on new developments in NLP.
• Constructed a Chatbot for Divar's Tech support agents, reducing the need for manual ticket re-spending by 50% which would lead to a dramatic decrease in workload and response time and increase accuracy.
• Participated in the interview process, contributing to team growth and talent acquisition

Technologies: Python, Pandas, Jupyter, BERT variants, GPT2, PyTorch

Theory: Transformers, Clustering, Chatbots

Divar, the largest classified Ad platform in Iran, with over 20 million (at the time) active users

• Query Recommendation System: Lead the development of a query recommendation system to address low-result queries. We analyzed millions of user action chains to identify co-occurring queries within a single session and using a W2V-like embedding calculated a query-query similarity metric. This strategic approach resulted in a substantial reduction of the search bounce rate for these problematic queries by over 50%.
• Fraud Detection System: Built upon preliminary work, advancing a fraud detection system in its development phase. Navigated strict data access limitations, necessitating justification for each data field required. Did extensive data cleaning and extraction. Implemented advanced clustering techniques to identify suspicious ad patterns and developed algorithms to detect broader networks of fraudulent accounts, laying the groundwork for enhanced platform security
• Search Enhancement: Led optimization efforts, including data annotation for supervised model training, automatic spelling correction, category recommendations based on user activity, and introducing similar ad suggestions to reduce bounce rates and enhance user experience. Liaison with Product Management: Worked closely with the Product Manager, monitoring metric changes and explaining their effects during A/B tests for new features. Also responsible for turning technical data into understandable business insights to help make informed decisions
• Led and Enhanced Recruitment Initiatives: Collaborated with a team of 4 Data Scientists, to overhaul the interview process, formulating the majority of interview questions. Screened over resumes, conducted interviews for 70+ candidates, and evaluated more than 20 projects to ensure the selection of top-tier talent

Technologies: Python, Pyspark, SQL, Apache Zeppelin, Airflow, Grafana, Docker, Kafka, fasttext

Theory: Search Engine, Ranking, Recommendation, Clustering, Fraud Detection

Cafe Bazaar, the largest Android marketplace in Iran, serves over 40 million active users, offering Android applications and games.

• App-Query Similarity: Collaborated with the lead data scientist to develop a Page-Rank inspired solution for identifying apps related to specific queries based on user interactions. Combined this with our existing text-based approach, resulting in a significant 8% reduction in search bounce and an improvement in our ranking metric
• Natural Language Toolkit: Developed a comprehensive toolkit for Persian language processing due to the absence of flexible existing tools. The toolkit included normalizers, tokenizers, lemmatizers,a POS tagger, and a spelling correction feature to cater to our specific needs.
• Metrics and Reporting: As the sole data scientist on the team, I worked closely with the Product Manager. My responsibilities included defining and reporting on key metrics, as well as translating metrics into clear business insights to guide decision-making.
• Interviewing at Bazaar: Joined the interview team, participating in 20+ interviews. Started as a secondary interviewer and progressed to leading interviews, showcasing my growing expertise

Technologies: Python, Elastic Search, Casandra, MongoDB, Docker, Pyspark, SQL, Regex, R, NLTK

Theory: Search Engine, Ranking, NLP