Thesis: Cost-Aware Tiered Website Interaction with Calibrated Decision Models

Growing use of AI agents on websites creates new demands on operating costs and infrastructure. This master’s thesis investigates whether a website assistant can reduce token consumption, cost and response time without compromising task completion by routing requests between lightweight decision models and larger language models. The project is carried out in collaboration with byBrick Interface, with particular attention to whether model confidence provides a reliable basis for escalation.

The assignment

• Review relevant research on tiered assistants, model confidence and escalation decisions.
• Design and implement a prototype for website tasks such as product discovery, search filtering, order lookup and review submission.
• Use structured website actions when appropriate and escalate to larger language models when confidence is too low or language generation is required.
• Evaluate the prototype against LLM-only approaches using representative website requests, measuring task success, token consumption, cost and latency, including routing overhead.
• Produce a reproducible evaluation suite and practical recommendations for balancing reliability, speed and operating costs.

Who we are looking for

• Master’s students in Computer Science, Software Engineering, Artificial Intelligence, Machine Learning or a related discipline, ideally applying as a pair.
• Strong programming ability, for example in Python or JavaScript/TypeScript, alongside an interest in web applications and APIs.
• Basic machine learning knowledge and an interest in LLMs, AI agents and experimental evaluation.
• Ability to work independently, assess results critically and present findings clearly.
• Experience with Docker or model APIs is an advantage.
• Expected to graduate in June 2027.
• Fluent spoken and written English. Swedish fluency is normally expected, but applicants without it are welcome to apply for individual assessment.

What we offer

• A relevant use case supplied by byBrick Interface, together with regular feedback during the project.
• An opportunity to develop and evaluate a practical AI website assistant in collaboration with industry.

Practical details

Based at the Västerås office, starting in early January 2027. Applicants are preferably working in pairs. Apply through the careers page as soon as possible, no later than 30 November 2026. Questions can be directed to Aldin Berisa or Edin Jelacic.