The BAPPEDA portal of Pariaman City previously served planning documents only as downloadable files, stored public complaints without processing, and offered no socioeconomic indicator data. During an internship at the Communication and Informatics Office of Pariaman City (1–31 July 2026), three artificial intelligence models were integrated into the portal through incremental development. First, a Retrieval-Augmented Generation (RAG) question-answering model combines gemini-embedding-001 vectors, cosine similarity, a similarity threshold, and Gemini 2.5 Flash with schema-constrained JSON output. Second, a zero-shot Large Language Model (LLM) analyses each complaint in the background and recommends category fit, sentiment, and urgency without sending the complainant’s identity. Third, linear regression, ARIMA, and Prophet forecast the Human Development Index (HDI), poverty rate, open unemployment rate, and Gini ratio of Pariaman City; the best method per indicator is selected by backtesting, with linear regression as a fallback. An evaluation with 40 test questions moved the threshold from 0.65 to 0.74, which gave the highest balanced accuracy (62.9%) and raised out-of-scope rejection from 14.3% to 64.3%, while recall@5 fell from 92.3% to 61.5%. Backtesting showed that linear regression was most accurate for HDI (MAPE 0.44%), whereas ARIMA was best for the other three indicators (MAPE 3.80–8.19%). All 18 automated feature tests passed. The models turn the portal into an interactive service that supports information retrieval, complaint prioritisation, and development planning.
Perum BULOG Bukittinggi Branch Office still monitors Sales Order (SO) transactions, records daily stock allocations, and identifies Rumah Pangan Kita (RPK) partners at risk of churn manually through separate Excel files, which makes recapitulation slow and allocation priorities subjective. This study aims to build a web-based dashboard that integrates SO quantity recapitulation, allocation tracking, and two partner classification models. The dashboard was developed using Python and Streamlit through seven stages: observation, data collection, requirement analysis, design, implementation, testing, and documentation. Model 1 classifies partner activity using Recency and Frequency (RFM), while Model 2 determines allocation priority using a Two-Stage Hybrid approach (RFM and Decision Tree) that is compared with Simple Additive Weighting (SAW). Of 288 partners, 152 (52.8%) were active, 6 (2.1%) were at risk, and 130 (45.1%) were passive or archived. SAW distinguished priority levels more finely than the Two-Stage Hybrid, whereas the Two-Stage Hybrid was simpler and easier to trace. Black box testing of 19 scenarios initially showed 16 valid scenarios (84.2%); the three failures, related to data type validation and partner identification, were fixed by validating input types and using buyer ID as the unique key, and were valid on retesting. This dashboard has the potential to support data-driven decision-making for stock distribution; however, its benefits have not yet been quantified through time-based metrics or user evaluations comparing processes before and after implementation.
This study develops an automatic document validation feature for a web-based testing application used by UPTD BMSPP, West Sumatra, to process agricultural machinery testing applications. Previously, the system checked uploaded files only by format and size, not by content, so staff inspected every document manually and mismatched documents were detected only after processing. Here, “validation” means checking whether a document’s text matches its expected document type, it does not verify authenticity or legal validity. The study builds a model combining Optical Character Recognition (OCR) using the Tesseract engine with rule-based keyword matching for five document types. The model runs as a separate Python Flask microservice: it extracts text from digital PDFs using pdfplumber, applies Tesseract OCR to scanned documents and images, then compares the text with per-type keyword lists to return a real-time status. The system was developed using the Waterfall method and tested with black-box testing on six scenarios. Five of the six scenarios (83%) produced the expected results, this shows scenario conformity, not OCR recognition accuracy or classification accuracy on a broad document population. One scenario returned an error message that did not match the expected response for an unsupported file format, marking an area for improvement.
Quality control in the chemical industry requires accurate and timely recording of process parameters. At PT Multi Nitrotama Kimia, laboratory results from three production plants—the Ammonium Nitrate Plant, Nitric Acid Plant, and Utility Plant—were recorded on paper logsheets and compiled in separate spreadsheets. This practice increased transcription errors, delayed report circulation, and provided no automatic warnings for parameters exceeding specified limits. This study aims to develop ANLAB, a web-based Laboratory Analysis Information System. The system was developed using the Waterfall model, PHP, CodeIgniter 3, and MySQL. Data were collected through observations, interviews with analysts, Distributed Control System (DCS) operators and supervisors, and document reviews. ANLAB features digital forms that classify results as normal, above, or below limits; a three-tier approval workflow; additional check requests; plant shutdown tagging; role-based access control; and PDF and Excel report exports. Black box testing involving 27 scenarios confirmed that all tested functions produced the expected results. However, testing focused exclusively on functionality; time savings, error reduction, and paperless performance were not measured. Future evaluations should assess these indicators using system timestamps. The findings indicate that ANLAB provides a structured and traceable digital workflow for laboratory data management in chemical plant operations
Rendahnya kemampuan alarm konvensional dalam memastikan pengguna benar-benar terbangun masih menjadi penyebab keterlambatan bangun pada sebagian pelajar, mahasiswa maupun pekerja. Sebagian besar aplikasi alarm konvensional hanya mengandalkan keluaran berupa suara atau getaran tanpa mekanisme untuk memastikan bahwa pengguna benar-benar telah terbangun. Penelitian ini bertujuan merancang aplikasi WakeUp AI, yaitu aplikasi alarm cerdas berbasis Android yang memanfaatkan kecerdasan buatan untuk memverifikasi kondisi pengguna sebelum alarm dapat dihentikan. Sistem menerapkan metode Transfer Learning dengan arsitektur MobileNetV2 untuk mengklasifikasikan kondisi mata terbuka dan mata tertutup. Sebelum proses klasifikasi dilakukan, citra dipraproses menggunakan teknik Eye Region Cropping sehingga model hanya memproses area mata. Selanjutnya model dijalankan secara lokal menggunakan TensorFlow Lite agar dapat bekerja tanpa koneksi internet. Untuk meningkatkan kestabilan hasil deteksi diterapkan algoritma Rolling Average dan Leaky Bucket Accumulator, kemudian proses verifikasi dilanjutkan dengan mengimplementasikan tantangan Random Question Challenge sederhana sebagai verifikasi kognitif pengguna. Berdasarkan implementasi dan pengujian yang telah dilakukan, sistem mampu menghasilkan aplikasi alarm yang ringan, cepat, serta lebih efektif dibandingkan alarm konvensional dalam memastikan pengguna benar-benar berada pada kondisi sadar sebelum alarm dinonaktifkan.
