Projects

Applied-AI systems built for platform, product, and research teams, followed by selected public work with source.

Where
Bentara Fraktal Indonesia Technologies
When
2025–2026
Role
AI Full Stack Engineer
Stack
Toba language models, Custom KV cache, Concurrency testing

Making small language models keep up under load

Toba is a family of in-house trilingual language models. I engineered a custom KV cache layer for their inference path, then validated it under heavy concurrent load to make sure the gains held.

Outcome Throughput held up under concurrent load testing.

Where
Bentara Fraktal Indonesia Technologies
When
2025–2026
Role
AI Full Stack Engineer
Stack
Multi-step agent workflow, Retrieval-augmented generation, Payment gateway, Model inference on bare metal, RTX 4090 cluster, Apple M2 Ultra

AI agents for a cultural library, from retrieval to bare metal

I built and deployed BFI’s cultural AI agents end to end. Each request runs as a series of steps, from retrieval-augmented generation over the Perpustakaan Digital Budaya Indonesia database to the payment gateway integration.

I also managed model inference across the bare-metal production setup: a multi-GPU RTX 4090 cluster and an Apple M2 Ultra node.

Outcome In production on bare-metal hardware, mixing NVIDIA GPUs and Apple silicon.

Where
Bentara Fraktal Indonesia Technologies
When
2025–2026
Role
AI Full Stack Engineer
Stack
Distributed web crawling, LLM-based generation

A knowledge corpus on Indonesian culture, built at scale

I architected an automated content pipeline that crawls the web and generates articles on Indonesian culture (Budaya Indonesia). Distributed web crawling feeds LLM-based generation to build a high-volume knowledge corpus.

Outcome An automated pipeline turns crawled pages into articles on Indonesian culture.

Where
Pukul Enam
When
2025
Role
Machine Learning Engineer
Stack
YOLOv8, Roboflow, NVIDIA Jetson Nano

Real-time poultry monitoring on edge devices

For an automated poultry monitoring system, I built the pipeline that turns camera feeds into annotation-ready images, and led a team of annotators doing polygon segmentation in Roboflow on crowded frames.

I also designed the object detection and inference pipeline: YOLOv8 for multi-camera observation, running on NVIDIA Jetson Nano.

Outcome Real-time, multi-camera tracking and analysis on the edge.

Where
Apple Developer Academy @ Infinite Learning
When
2024
Role
iOS Developer, AI/ML Engineer
Stack
Apple Speech framework, Gemini Pro, Google text-to-speech

Indonesian speech-to-speech on iOS

I engineered an Indonesian speech-to-speech system that combines Apple’s native Speech framework, Gemini Pro, and Google text-to-speech.

Optimized data structures and custom text normalization shortened the wait between speaking and hearing a reply.

Outcome A shorter wait between speaking and hearing the reply.

Where
Data Science and Artificial Intelligence Lab, Universitas Sumatera Utara
When
2023
Role
Technical Research Assistant
Stack
TensorFlow, Distributed training, Flask

Reading Batak script for a learning platform

I joined a research collaboration between Universitas Sumatera Utara and Leiden University as the AI/ML engineer.

I designed the OCR pipeline for a Batak script learning platform in TensorFlow, trained it on a multi-GPU server, and served the model with Flask from the department server for students in remote areas.

Outcome The model runs on the department server and serves students in remote areas.

Public code I've written.

  1. MacMultipeer Bundle (opens in new tab)

    2024 native systems

    a Swift bridge that exposes encrypted Apple Multipeer Connectivity sessions to Unity on macOS.

    local peer discovery · reliable messaging · no server

    • Swift
    • Unity
    • MultipeerConnectivity

For roles, teams, and the longer timeline, see about me.