AI engineer · applied machine learning

Widi Nugroho

Data scientist and AI engineer. I build machine learning systems across audio, vision and retrieval — and I write down what the numbers actually say.

M.Mat. in Mathematics, research track · Valedictorian, GPA 4.00/4.00 · Universitas Indonesia

Widi Nugroho in academic dress, seated on the brick steps of his graduation venue, holding his cap.
Focus: applied machine learning — audio, computer vision and retrieval.

Measured · as reported in each paper

86.5%

accuracy for smoker status from raw cough audio — Wav2Vec2 over 7,561 smoker and 7,561 non-smoker samples. F1 0.862 · AUC 0.945.

90%

accuracy detecting tuberculosis from cough audio plus clinical metadata, on an unseen test set of 16,000 balanced samples. Precision, recall, specificity and F1 all 0.90.

5

peer-reviewed papers, 2025–2026 — two IJACSA journal articles, one MDPI Sensors article and two IEEE conference proceedings, from cough acoustics to respiratory rate from video.

Peer-reviewed · 2025–2026

Five peer-reviewed papers across audio, vision and clinical data.

Every result re-run under a subject-level split, with no leakage between train and test — and where a previously published claim did not survive that correction, it is written down here too.

Journal article IJACSA · Vol. 16, No. 9 · 2025

Performance analysis of spectrogram-based versus raw-waveform deep learning models for smoker detection from cough audio

Widi Nugroho, Alhadi Bustamam, Rinaldi Anwar Buyung

The spectrogram model's accuracy was reported as a range; the bar marks the midpoint. 7,561 smoker and 7,561 non-smoker samples from the CODA TB DREAM Challenge.

PythonPyTorchWav2Vec2AST

10.14569/IJACSA.2025.0160948

Journal article IJACSA · Vol. 16, No. 10 · 2025

Multimodal deep learning for tuberculosis detection using cough audio and clinical data with Health Acoustic Representations (HeAR)

Rinaldi Anwar Buyung, Widi Nugroho

Accuracy on the unseen test set. Precision, recall, specificity and F1 all 0.90 on both classes — the paper reports them at one digit of precision, so no per-class spread is shown.

PythonPyTorchHeARMultimodal fusion

10.14569/IJACSA.2025.0161036

Conference proceeding ICMCTC · 2025

Smoker detection from cough sounds: a deep learning approach using AST

Widi Nugroho, Alhadi Bustamam, Rinaldi Anwar

Non-invasive smoking-status classification from cough audio, with attention pooling over the spectrogram patches.

PythonPyTorchASTAttention pooling

10.1109/ICMCTC62214.2025.11196138

Conference proceeding ICCRAIDS · 2026

Efficient rPPG via hybrid SSM-attention: a global context, periodic focus approach

Rinaldi Anwar Buyung, Alhadi Bustamam, Hendri Murfi, Harum Ananda Setyawan, Widi Nugroho

PulseMamba is an architecture contribution: the paper reports design and linear cost, not a benchmark score, so no accuracy bar is drawn for it.

PythonPyTorchrPPGState space models

10.1109/ICCRAIDS67816.2026.11519582

Journal article Sensors (MDPI) · Vol. 26, No. 22 · 2026

Colour-channel selection for video-based respiratory rate estimation: a diagnostic study with a single-shot held-out evaluation

Rinaldi Anwar Buyung, Widi Nugroho, Harum Ananda Setyawan, Muhammad Remzy Syah Ramazhan, Ahmad Ihsan Farhani

Lower is better: mean absolute error in breaths/min, all four bars on one scale so a shorter bar means a smaller error. The facial channels are reported as a range; the bar marks 3.69. COHFACE, 39 subjects and 160 clips, seated participants, subject-disjoint splits with the held-out rows scored once.

Trained two-stream fusion sits at 3.439; picking a channel per window brings it to 2.463 — still behind a last-measurement baseline of 1.438, and not distinguishable from a blind guess (p = 0.232). Removing the motion component does not hurt the colour estimate (1.156), so the mechanism is narrowed rather than explained. The paper's verdict is measurement-first: region and channel choice dominate architecture on these data, and a single error number means nothing without a same-row baseline, clinical-threshold shares and per-subject spread.

PythonrPPGSignal processingCOHFACE

10.3390/s26227107

Measured performance, exactly as reported

Study Model Accuracy F1 AUC
Smoker detection · IJACSA 2025 Wav2Vec2 · raw waveform 86.5% 0.862 0.945
Smoker detection · IJACSA 2025 AST · spectrogram, baseline ≈76–77% — —
Tuberculosis · IJACSA 2025 HeAR + clinical, late fusion 90.0% 0.90 —
Smoker detection · ICMCTC 2025 AST + attention pooling 81.62% — —
rPPG · ICCRAIDS 2026 PulseMamba · hybrid SSM-attention not reported — —
Respiratory rate · Sensors 2026 Untrained torso chrominance — · MAE 1.267 — —

A dash means the paper does not report that metric. Sample counts: 7,561 smoker and 7,561 non-smoker cough samples in the CODA TB DREAM Challenge set, and 16,000 balanced samples for the multimodal model. PulseMamba is an architecture contribution, reported here as published rather than with a score it does not claim. The respiratory-rate study is regression, not classification, so its column reads as mean absolute error in breaths/min against a 2.644 same-row blind baseline.

Project · retrieval-augmented generation

ask-the-docs — answers that show where they came from.

A question about a documentation corpus is answered only from the sections retrieved for that question, naming the section each claim came from. No framework sits in the pipeline — chunking, index, retrieval and the model client are all written out by hand.

  1. 01

    Heading-aware chunking

    Splits the corpus on headings and never inside a code block, so a snippet reaches the model intact.

  2. 02

    Local CPU index

    sentence-transformers embeddings in Chroma, served from CPU — the retrieval layer costs nothing to keep warm.

  3. 03

    FastAPI behind Docker

    One provider-agnostic model client, so the model behind the answer can change without touching the pipeline.

  4. 04

    The sources panel

    Every answer shows which passages were actually sent to the model and how close each one was — the part that makes it auditable.

6/12

retrieval hit@5 on a fixed 12-question set

0.500

MRR — mean reciprocal rank of the first correct passage

6/12

answers that name the section their claim came from

3.9 s

p50 latency, warm container

Built with

PythonFastAPIDockerChromasentence-transformersOllama API

Open it

Live · running now

Ask it something.

Type a question about the Ollama documentation, then open the sources panel under the answer — that panel is the part that makes the result checkable instead of taken on trust. The container sleeps when idle, so the first load can take a few seconds.

Demo password portofolio_widinugroho Open the demo in a new tab

Waking the demo container — the first load can take a few seconds while the CPU index comes up.

Measured on a fixed 12-question set · API-reference pages exist in the corpus only as OpenAPI YAML, so questions aimed at them cannot be answered from it · uploaded PDFs are not measured at all.

Experience

Where the work happens.

Real work arrives as messy data — a recorded cough, a phone camera, a scanned bank statement. Most of the value is in the evaluation protocol, not the architecture.

  1. Sep 2024 – Present

    Data Scientist

    PT Seleris Meditekno Internasional

    • Low-computational-cost rPPG app for heart rate and HRV estimation, built for insurance underwriting.
    • Research on blood pressure and hemoglobin estimation from facial video.
    • OCR pipeline for bank statements.
    • Cough-sound models for tuberculosis detection and smoker status.
  2. Oct 2023 – Sep 2024

    Research Assistant

    Data Science Center FMIPA UI

    • Grant proposal writing and intellectual-property administration.
    • Diabetes detection prototype from fundus images.
    • Taught data science to PT Pamapersada Nusantara.
    • Technical committee member for conferences.
  3. Jul – Oct 2023

    Data Analyst Intern

    PT Bank Permata Tbk

    • Monthly reporting in Excel.
    • Process automation with VBA macros.
  4. Dec 2022

    Full Stack Developer

    Order Management System · freelance project for a family business

    • Web app that turns bedsheet orders pasted from messy WhatsApp templates into structured database records.
    • Stack: Kotlin, Ktor, Vaadin, and Regex for parsing the incoming messages.
  5. Jul 2022

    Teaching Assistant

    Data science training for PT Telkom Indonesia

    • Wrote the training material for data processing with pandas, PyTorch and time series.
    • Delivered the programme alongside the lead trainer.

Education & skills

Formal training in mathematics, applied to machine learning.

Education

Aug 2024 – Aug 2026

M.Mat. in Mathematics, research track

Universitas Indonesia · specialization in Data Science

Valedictorian, GPA 4.00/4.00.

Aug 2019 – Aug 2023

S.Si. in Mathematics

Universitas Indonesia

GPA 3.44/4.00.

2022

Certifications

Data Scientist (Kampus Merdeka) — MyEduSolve. ITS Database and ITS Python — Certiport.

Skills

Languages & data

PythonGitExcelSQL (basic)

ML & AI

Deep learningComputer visionAudio & speechNLPOCR

Research

Scientific writingGrant proposalsPublications

Contact

Take the CV, or write to me.

The PDF covers education, roles and skills. The papers, their DOIs and the evaluation protocol behind each number stay on this page. Email is the fastest way to reach me.