Projects

This project is in colaboration with Logpilot, Sweden. Logpilot AI leverages proprietary predictive models and fine-tuned LLMs (Mistral AI) to anticipate and resolve issues before they disrupt your operations. We aim to develop a domain-specific AI automation agent that balances cost-effectiveness, scalability, and security. Students will fine-tune a compact open-weight language model (e.g., Mistral-7B or Llama 3-8B) using parameter-efficient techniques like LoRA or QLoRA. The agent will be integrated into an autonomous framework with memory and tool-use capabilities, inspired by architectures such as Reflexion and AutoGen. Emphasis will be placed on deploying the agent in a secure environment, ensuring compliance with data protection standards.

User-friendly apartment search with natural language queries. Built preference parsing system, intelligent ranking algorithms, and application guidance features to simplify the housing search experience.

Co-architected and developed a scalable agentic AI platform designed for production environments. Includes declarative agent specifications with YAML-based configuration, multi-tier memory systems (working, episodic, semantic), integration with vector stores and relational databases, comprehensive evaluation pipelines for agent testing.

Built an intelligent assistant for 500+ Linux shortcuts using RAG pipeline. Engineered vector embeddings and generative AI to answer natural language queries from man pages with high accuracy. Demonstrates practical application of retrieval-augmented generation for documentation search.

Implemented a fully functional Transformer architecture from scratch in PyTorch. Built multi-head self-attention, positional encoding, and encoder-decoder components. Benchmarked on sequence tasks and evaluated the impact of hyperparameter tuning on model performance.

Built distributed ML pipeline for sentiment analysis on 14M+ song lyrics, achieving 92% accuracy and 40% faster processing. Leveraged Hadoop ecosystem with Spark, PySpark, MapReduce, and HDFS for scalable data processing. Demonstrates expertise in big data engineering and distributed computing.