Large Language Models (LLMs) have demonstrated remarkable generative capabilities
across diverse domains, yet their effectiveness in multi-tool, multi-domain applications re-
mains constrained by the limitations of single-shot prompting. This paper presents Merid-
ian, a modular conversational architecture where a single core LLM engine (Google Gemini)
serves nine domain-specific tools—each governed by its own system prompt—through a con-
versational interface that collects clarifying context from users prior to generation. We in-
vestigate whether conversational context-gathering before generation improves LLM output
quality compared to zero-shot approaches. Through preliminary qualitative evaluation with
12 users across three use cases (CV building, SOP writing, and study plan generation), we
observe that 2–3 rounds of clarifying questions increase output relevance and substantially
reduce hallucination rates. We further conduct a systematic limitation analysis identifying
key failure modes including context window constraints, prompt sensitivity, and evalua-
tion challenges. This work contributes (1) a modular architecture for multi-domain LLM
applications deployable on free-tier infrastructure, (2) empirical evidence for the value of
conversational context-gathering, and (3) a framework for rigorous limitation analysis in
applied LLM research. The architecture is fully open-source and deployed on Streamlit
Community Cloud.
Read the Full paper (https://github.com/sultanalikhan7543/Meridian/blob/main/Merdian_research_paper.pdf)
Top comments (0)