Applied Research Laboratory

Interpreting dental radiographs requires meticulous clinical observation, with diagnostic consistency heavily reliant on practitioner experience. This applied research investigates deep learning computer vision models capable of identifying and highlighting suspected pathological anomalies to assist specialist review.

The initiative originated directly from industry-academia teaching within the Digital Dentistry and Automated Detection course at Universitas Gadjah Mada’s Dental Radiology Specialist Program. This partnership ensures that algorithmic development is directed by actual clinical workflows, medical imaging standards, and diagnostic criteria rather than isolated computational benchmarks.

Our clinical boundary is uncompromising: AI serves strictly as a decision-support assistant rather than an autonomous diagnostic agent. Any transition toward clinical trial or deployment is predicated upon rigorous clinical validation, ethical oversight, and clear professional accountability.

Users of High-Performance Computing (HPC) clusters frequently encounter repetitive hurdles: job scheduling failures, module configuration conflicts, and runtime environment issues. This research developed a targeted language model prototype to resolve these first-tier inquiries reliably.

The project originated as an M.Sc. Data Science thesis at EURECOM (France), complemented by practical field implementation at EFISON Lisan Teknologi.

The enduring takeaway from this initiative is methodological rather than purely algorithmic: model sizing and hardware provisioning are direct cost drivers. Sound technical choices must be grounded in rigorous empirical benchmarking, not speculation.

The vast majority of standardized Large Language Model (LLM) benchmarks were engineered for and calibrated against English corpora. When these models are deployed across Indonesian institutional environments—characterized by specialized legal phrasing, bureaucratic nomenclatures, and regional code-mixing—standard global benchmarks fail to accurately predict real-world reliability.

This research develops domain-specific evaluation protocols tailored to the Indonesian linguistic landscape: quantifying hallucination rates when queries exceed source context boundaries, verifying attribution accuracy and citation integrity, and assessing semantic reasoning across mixed-register documents.

We actively welcome research collaboration, particularly with university computational linguistics laboratories and academic teams maintaining indigenous and regional language corpora.

A significant share of Indonesian enterprise workflows and public administrative processes remains tethered to document artifacts: identity credentials, tax invoices, transaction vouchers, insurance claim files, and regulatory records. Converting these unstructured records into queryable, structured databases yields massive operational efficiencies, yet remains technically challenging due to non-standard layouts, noise, and uneven scan resolutions.

This research builds directly upon our text engineering and natural language processing (NLP) methodologies developed during our advisory work with BPIW at the Ministry of Public Works and Housing. Our core research focus centers on adaptive confidence calibration: engineering models that reliably evaluate their own extraction certainty and automatically route ambiguous fields to human operators for verification.

We document this initiative as an active research track rather than an off-the-shelf commercial product. We welcome applied research partnerships with enterprise and public sector organizations managing large-scale document archives.