I'm Mahan Dehghani, a bachelor's student in Computer Science at K. N. Toosi University of Technology (KNTU).
I build production-ready applications at the intersection of Retrieval-Augmented Generation (RAG) and agentic AI, with a broader focus on applications powered by large language models. I enjoy connecting models to useful knowledge, tools, and reliable software systems.
My research interests lie at the intersection of bioinformatics and LLM interpretability: understanding how language models represent and reason about biological information.
- RAG applications: document ingestion, chunking, embeddings, vector search, hybrid retrieval, reranking, and answers grounded in cited sources.
- Agentic workflows: tool calling, structured outputs, stateful orchestration, memory, and human-in-the-loop interactions.
- LLM-powered products: conversational assistants, knowledge search, and APIs that connect models to application data and tools.
- Production foundations: asynchronous APIs, input validation, automated tests, containerized deployment, evaluation, tracing, and attention to latency and cost.
My stack and the ecosystem I continue to explore — from retrieval experiments to deployed applications.
Beyond frameworks, I focus on prompt design, tokenization, embeddings, context management, semantic search, BM25, reranking, JSON Schema, REST APIs, SQL, Linux, and CI/CD. For model adaptation and serving, the ecosystem also includes PEFT/LoRA, quantization, and vLLM.
I'm interested in questions where bioinformatics meets interpretable language models:
- How do language models encode biological sequences and scientific knowledge?
- What can probing, attribution, and mechanistic interpretability reveal about their internal representations?
- How can retrieval over biological literature support traceable, evidence-grounded scientific reasoning?
- How can we evaluate biological relevance, uncertainty, and faithfulness in model outputs?
I want to build a deeper understanding of both the data and the models that learn from it.
From retrieved knowledge to useful agents — with curiosity about the biology behind the data.
mahandhgh :: retrieve → reason → build → understand