Arya Prasetya, machine learning engineer, Bandung

Clusters found, not guessed

Scroll and the pieces of data above find their clusters, one for each system I have built. The drawing is an illustration and the order is real: the model forms the clusters, a human names them. That is how nearly all of my work begins.

Talk about your project See projects

Latency and cost written downEvaluation before launchCode handed over

Arya Prasetya

Machine learning engineer

Bandung, West JavaTaking one new project from April 2026

Get in touch

Eight years, the countable part

8 years years building machine learning systems

14 systems still running in production

180 ms millisecond p95 on the busiest system

31 percent inference cost cut after quantisation

Three systems, written constraint first

Complaint clustering, an insurer

Nine thousand complaints a month across four channels, read one by one. The constraint was not accuracy: the team would not trust a label they could not check.

Clustering with no starting labels, then every cluster named with the team. Reading time fell from six hours to forty minutes a day, and every decision traces to its three nearest examples.

Embeddingsclustering2025

Policy document search, a state enterprise

Twenty two thousand pages of overlapping internal rules. Keyword search failed because each directorate used different terms.

Semantic search with mandatory citations: every answer points at the original page. p95 of 180 milliseconds on hardware they already owned, with no GPU.

Semantic searchon premise2024

Defect detection, a components factory

Twelve thousand parts inspected by eye per shift, with attention falling sharply after two hours.

A small model on a single board computer on the line, 40 milliseconds per part. Escaped defects fell 64 percent, and the operator still decides.

Computer visionedge2023

Open to the public

embed-id A small Indonesian embedding model, retrained from an open model for customer service text. Python1.2k retrieval-eval A search evaluation tool that makes you write the test questions before you see any results. Python410 tflite-jalur A full example of running a vision model on a single board computer on a production line, latency measurements included. C++260

What I do myself

Models

Embeddings and retrievalText classificationVision for inspectionFine tuning open models

Production

Serving to a latency budgetQuantisation and distillationOutput quality monitoringA and B testing

Tools

PythonPyTorchONNXTensorRTTFLitePostgres with pgvectorDockerKubernetes

What I do not do

Training large models from scratchProjects with no test setSystems with no human in the loop

Those four columns, including the one left uncast

Eight years

2021 to now
Freelance engineer

Independent, Bandung

Working directly with product teams, usually three to six months per system, until it runs without me.
2019 to 2021
Senior engineer

A logistics company

Demand forecasting and routing. Here I learned that a model two percent better but hard to explain gets abandoned.
2017 to 2019
Data scientist

A fintech startup

Risk scoring and fraud detection on data far dirtier than anything university had shown me.

Writing

Why semantic search fails on internal documentsPersonal blog, 2025
Eight bit quantisation for small Indonesian language modelsNational Computer Science Conference, 2024
Evaluation before launch: a nine point checklistPersonal blog, 2023

Education and certification

MSc Informatics

Bandung Institute of Technology

2017

BEng Electrical Engineering