Stefanus Kurniawan

Stefanus Kurniawan

Auto-generated profile, based on publicly available information

Stefanus Kurniawan

Applied Artificial Intelligence Initiative, Deakin University, Geelong, Australia

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
14 April 2024
Verified
Conference Paper
Open access
SBScott Barnett
Scott Barnett
SKStefanus Kurniawan
Stefanus Kurniawan
STSrikanth Thudumu
Srikanth Thudumu
ZBZach Brannelly
Zach Brannelly
MAMohamed Almorsy Abdelrazek
Mohamed Almorsy Abdelrazek
Software engineers are increasingly adding semantic search capabilities to applications using a strategy known as Retrieval Augmented Generation (RAG). A RAG system involves finding documents that semantically match a query and then passing the documents to a large language model (LLM) such as ChatGPT to extract the right answer using an LLM. RAG systems aim to: a) reduce the problem of hallucinated responses from LLMs, b) link sources/references to generated responses, and c) remove the need for annotating documents with meta-data. However, RAG systems suffer from limitations inherent to information retrieval systems and from reliance on LLMs. In this paper, we present an experience report on the failure points of RAG systems from three case studies from separate domains: research, education, and biomedical. We share the lessons learned and present 7 failure points to consider when designing a RAG system. The two key takeaways arising from our work are: 1) validation of a RAG system is only feasible during operation, and 2) the robustness of a RAG system evolves rather than designed in at the start. We conclude with a list of potential research directions on RAG systems for the software engineering community.

14 January 2025 in Journal of Medical Internet Research
Verified
Journal Article
Open access
WZWu Yi Zheng
Wu Yi Zheng
ASArtur Shvetcov
Artur Shvetcov
ASAimy Slade
Aimy Slade
ZJZoe Jenkins
Zoe Jenkins
LHLeonard Hoon
Leonard Hoon
JNJill M Newby
Jill M Newby
Background: With increasing adoption of remote clinical trials in digital mental health, identifying cost-effective and time-efficient recruitment methodologies is crucial for the success of such trials. Evidence on whether web-based recruitment methods are more effective than traditional methods such as newspapers, media, or flyers is inconsistent. Here we present insights from our experience recruiting tertiary education students for a digital mental health artificial intelligence–driven adaptive trial—Vibe Up. Objective: We evaluated the effectiveness of recruitment via Facebook and Instagram compared to traditional methods for a treatment trial and compared different recruitment methods’ retention rates. With recruitment coinciding with COVID-19 lockdowns across Australia, we also compared the cost-effectiveness of social media recruitment during and after lockdowns. Methods: Recruitment was completed for 2 pilot trials and 6 minitrials from June 2021 to May 2022. To recruit participants, paid social media advertising on Facebook and Instagram was used, alongside mailing lists of university networks and student organizations or services, media releases, announcements during classes and events, study posters or flyers on university campuses, and health professional networks. Recruitment data, including engagement metrics collected by Meta (Facebook and Instagram), advertising costs, and Qualtrics data on recruitment methods and survey completion rates, were analyzed using RStudio with R (version 3.6.3; R Foundation for Statistical Computing). Results: In total, 1314 eligible participants (aged 22.79, SD 4.71 years; 1079, 82.1% female) were recruited to 2 pilot trials and 6 minitrials. The vast majority were recruited via Facebook and Instagram advertising (n=1203; 92%). Pairwise comparisons revealed that the lead institution’s website was more effective in recruiting eligible participants than Facebook (z=3.47; P=.003) and Instagram (z=4.23; P<.001). No differences were found between recruitment methods in retaining participants at baseline, at midpoint, and at study completion. Wilcoxon tests found significant differences between lockdown (pilot 1 and pilot 2) and postlockdown (minitrials 1-6) on costs incurred per link click (lockdown: median Aus $0.35 [US $0.22], IQR Aus $0.27-$0.47 [US $0.17-$0.29]; postlockdown: median Aus $1.00 [US $0.62], IQR Aus $0.70-$1.47 [US $0.44-$0.92]; W=9087; P<.001) and the amount spent per hour to reach the target sample size (lockdown: median Aus $4.75 [US $2.95], IQR Aus $1.94-6.34 [US $1.22-$3.97]; postlockdown: median Aus $13.29 [US $8.26], IQR Aus $4.70-25.31 [US $2.95-$15.87]; W=16044; P<.001). Conclusions: Social media advertising via Facebook and Instagram was the most successful strategy for recruiting distressed tertiary students into this artificial intelligence–driven adaptive trial, providing evidence for the use of this recruitment method for this type of trial in digital mental health research. No recruitment method stood out in terms of participant retention. Perhaps a reflection of the added distress experienced by young people, social media recruitment during the COVID-19 lockdown period was more cost-effective. Trial Registration: Australian New Zealand Clinical Trials Registry ACTRN12621001092886; https://tinyurl.com/39f2pdmd; Australian New Zealand Clinical Trials Registry ACTRN12621001223820; https://tinyurl.com/bdhkvucv

14 April 2024
Verified
Conference Paper
Open access
ZRZafaryab Rasool
Zafaryab Rasool
SBScott Barnett
Scott Barnett
DWDavid Willie
David Willie
SKStefanus Kurniawan
Stefanus Kurniawan
SBSherwin Balugo
Sherwin Balugo
MAMohamed Almorsy Abdelrazek
Mohamed Almorsy Abdelrazek
Large language models (LLMs) enable state-of-the-art semantic capabilities to be added to software systems such as semantic search of unstructured documents and text generation. However, these models are computationally expensive. At scale, the cost of serving thousands of users increases massively affecting also user experience. To address this problem, semantic caches are used to check for answers to similar queries (that may have been phrased differently) without hitting the LLM service. Due to the nature of these semantic cache techniques that rely on query embeddings, there is a high chance of errors impacting user confidence in the system. Adopting semantic cache techniques usually requires testing the effectiveness of a semantic cache (accurate cache hits and misses) which requires a labelled test set of similar queries and responses which is often unavailable. In this paper, we present VaryGen, an approach for using LLMs for test input generation that produces similar questions from unstructured text documents. Our novel approach uses the reasoning capabilities of LLMs to 1) adapt queries to the domain, 2) synthesise subtle variations to queries, and 3) evaluate the synthesised test dataset. We evaluated our approach in the domain of a student question and answer system by qualitatively analysing 100 generated queries and result pairs, and conducting an empirical case study with an open source semantic cache. Our results show that query pairs satisfy human expectations of similarity and our generated data demonstrates failure cases of a semantic cache. Additionally, we also evaluate our approach on Qasper dataset. This work is an important first step into test input generation for semantic applications and presents considerations for practitioners when calibrating a semantic cache.

12 April 2026
Verified
Conference Paper
Open access
DWDilani Widanapathiranage
Dilani Widanapathiranage
SBScott Barnett
Scott Barnett
SKStefanus Kurniawan
Stefanus Kurniawan
WTWannita Takerngsaksiri
Wannita Takerngsaksiri
Hallucinations are a key concern when creating applications that rely on Foundation models (FMs). Understanding where and how these subtle failures occur in an application relies on evaluation methods known as evals. Prior work focuses on defining new eval methods or benchmark datasets for specific tasks. However, neither helps a software team with a task-specific FM application when there is no metric or dataset. The demand for both automated approaches and deep integration of human insight makes this a challenging problem. We address this gap by proposing an approach to synthesise a FM task-specific evaluator program that provides automation and a custom UI for capturing feedback. The core novelty of our approach lies in: (1) a task-agnostic meta-model that captures properties of any FM task, (2) an interaction protocol for efficient use of human feedback, and (3) an eval synthesiser that selects or generates an appropriate set of evals. We implement our approach in TaskEval and demonstrate the concept on two diverse FM tasks: chart data extraction and document question answering. A preliminary evaluation on the quality of our selected evals shows 93% and 90% accuracy respectively. Our research tackles a growing problem facing engineering teams, how to evaluate and review outputs from FM tasks.