Mirza K.-Agentic AI for a DeepResearch project

• This project involved extraction of intelligence data to support report writing for a company that provides geopolitical, global, commercial intelligence. The data have been gathered from a number of resources (interview transcripts, online data, internal documents), and then a knowledge base has been build from it. This was the basis of a complex RAG system, that was evaluated against a golden dataset. Agents have been used to find out the contradicting intelligence, the statements supporting each other, and to store back the generated knowledge.

Used: Python, RAG, LangGraph, LangChain, deepeval, MCP

• To create a multi-agentic system, supported by a Knowledge Graph, that automates the process of drafting a research paper. The system used multiple experts (OpenAI models) that 'collaborated' during the process of document drafting. The whole process was supported by a Knowledge Graph out of which we extracted useful information. Technologies used for this framework: LangChain, LangGraph, Smolagents, LlamaIndex, dspy. The project involved the use of Terraform and GitHub Actions (CI/CD Pipeline), via AWS. The initial application was deployed as a Streamlit app. I was responsible for the AI Engineering part, Knowledge Graph creation, and also deployment in AWS. This was a GenAI Engineer role, also involving Prompt Evaluation and Retrieval Optimization (Langfuse).

Used: Python, Git, LangChain, LangGraph, dspy, GraphRAG, RAG, SQL

• an LLM chatbot to help with HR-related inquiries. Approach: I led the development of an advanced Retrieval-Augmented Generation (RAG) system aimed at improving HR data retrieval processes. This system utilizes a Vector Similarity Search Database and commercial LLM API to efficiently source and integrate extensive HR-related data, enabling it to respond to a broad spectrum of HR inquiries. We have also attempted to fine-tune an LLM for this task, and the technologies used were PEFT (LoRA), and sophisticated Prompt Engineering. My responsibility was setting up the system, starting with plain HR documents, and using Amazon Bedrock for this project. We also associated several agents with the system.

Used: Python, Git, LangChain, LangGraph, Milvus, SQL

• Given a spectral image showing sensor reading values, classified the signals and identified novelties (anomalies).
• Since there was not enough labeled data, I relied on self-supervised machine learning methods. To meet the customer's need for fast processing, I used one of the YOLO architectures. The system was developed with the MMYOLO framework. I was responsible for the main modeling work.

Used: Python, Git, SQL

• Researched state-of-the-art developments in recommender systems for media (audio, video, and text content).
• Implemented a recommender system powering ARD Audiothek (one of Germany's most popular audio-on-demand platforms). The deployed production model had 15% higher precision than the previous one.
• NLP projects: entity recognition and redundancy removal

• To advise individuals on financial actions that enhance their chances of achieving specific goals.
• For an Asian bank aiming to offer actionable advice for long-term financial planning, such as securing a home purchase in 20 years.
• Predictive machine learning models were developed to suggest optimal actions (example: obtaining a salary increase of at least 10%).

• To surpass the performance of Amazon Personalize's Hierarchical Recurrent Neural Network-based recommendation system.
• Adopted a recommender model based on Factorization Machines, significantly improving recommendations through targeted feature engineering. Kafka was used for this project.

• To detect unusual behavior in a sensor-monitored network.
• Established a baseline machine learning model predicting expected sensor readings, using deviations from this model to flag anomalies. Enhanced detection accuracy with a Graph Neural Network, analyzing the connectivity between sensors and observable areas. The models used were GNNs supported in the PyG framework.

Used: Python, Git, SQL, PyG

• To predict vehicle transportation prices, providing dispatchers with a reliable basis for pricing.
• Enhanced the machine learning model's accuracy by approximately 40% through advanced feature engineering and implementing a two-tiered regression strategy, combining residual and standard regression techniques powered by XGBoost. The dispatcher portal grew notably in terms of user interaction with the board.

Used: Python, Git, SQL

• (re)implementation of a testing software
• implementation of Finite State Machines
• code generation