Prompt optimization method to maximize the accuracy and reliability of LLM outputs

Draft & Goal (D&G) is a Canadian artificial intelligence (AI) company developing a no-code platform that automates marketing workflows using AI agents for writing, search engine optimization, and fact-checking. The integration of large language models (LLMs) is central to their software solution: it allows non-technical users to harness advanced AI for content creation. However, unlike traditional software systems, LLMs are inherently non-deterministic and opaque, making their behavior difficult to predict or control.
Currently, D&G’s developers rely heavily on manual trial-and-error strategies to craft effective prompts (i.e., instructions input to the LLM) [1, 2]. This process is time-intensive, brittle, and often produces results that fail to generalize across domains or models [3]. The lack of systematic tools for prompt design leads to inefficiencies in development, higher operational costs, and limited scalability. Therefore, automating prompt optimization is critical for D&G’s ability to deliver consistent, high-quality outputs at scale.
Through this project, D&G aims to integrate data-driven methods for refining prompts, informed by execution feedback. The anticipated benefits include increased efficiency in product development, improved reliability of AI outputs, and a stronger competitive position in Canada’s growing AI software market. This project will also reduce costs and broaden accessibility of advanced AI tools for non-technical users.

Faculty Supervisor:

Eugene Syriani;Ian Arawjo

Student:

Partner:

Recherches Super Nova Inc.

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

Université de Montréal

Program: