Stepping away briefly from full-time engineering work to focus on an important life milestone.

• During this period, I spent some time working on personal hobby projects, including a home automation app and other small explorations, while keeping my technical skills active at a comfortable pace.

Preparing to transition back into full-time engineering work with fresh energy and perspective.

Served as an AI Engineer working directly for the CEO to research, design, and rapidly prototype AI-driven systems across multiple domains.

• February–May 2025

• Real-Time Audio & Multilingual Systems

• Built a real time multilingual audio translation system leveraging WebRTC and Meta Seamless model for low latency speech processing.

• Engineered an end to end pipeline for speech-to-text, translation, and text-to-speech including live transcription display and synthesized audio output.

• LLM Agents, RAG Systems & Knowledge Automation

• Built LLM agents for documentation, coding assistance, and transforming email content into structured knowledge.

• Designed a multi agent orchestration framework, real time streaming interface, and a vector-based KMS for long-context retrieval.

• Created an MVP developer assistant agent that learns from websites, codebases, and email Q&A to generate guides and summaries.

• Implemented RAG pipelines and delivered internal agents for JSON cleanup, recipe generation, and enhanced chatbot workflows.

• Marketing & Social Media Automation Agents

• Developed an AI engine that generates, refines, and schedules social posts, with conversational editing via local LLM.

• Built a firecrawl powered URL ingestion pipeline that converts webpages into post ready content with human-in-the-loop approval.

• Integrated Arcade API publishing, Slack notifications, LangSmith tracing, and delivered a containerized Postgres/Redis backed agent platform.

• Financial Automation

• Implemented automated workflows using the Interactive Brokers (IBKR) API for data retrieval and monitoring.

• Integrated existing financial logic into a modular and scalable containerized Python-based automation system.

• October–December 2024

• AI & Machine Learning Training

• Mentored and guided tutors on local LLM model training, fine-tuning, and implementation, providing hands-on support for real-world applications.

• Proof of Concept Development

• Designed and implemented a phone call voice-based authentication system, integrating voice recognition, biometric security, speech-to-text, text-to-speech and Twilio’s API for secure and scalable deployment.

Tools used: Python, JavaScript, PyTorch, CUDA, Scikit-learn, HuggingFace, FastAPI, Flask, Node.js,Ollama, LangChain, LangGraph, Firecrawl, Agent Inbox, LangSmith, Twilio, Postgres, Redis, Arcade API, Slack API, IBKR API, Pandas, NumPy.

Started in the R&D division conducting AI research and developing foundational LLM applications, and later advanced to AI Squad Lead where I oversaw ML/AI engineering execution, AI platform operations, and team collaboration.

• AI Research Engineer

• LLM Applications

• Initiated and led AI research projects on Large Language Models (LLM), starting with functionality in Retrieval-Augmented Generation (RAG) and real-time database integration. The project later became known as Mas Ahya

• Created, engineered, and led the breakthrough development of Mas Ahya (a generative-AI-powered conversational assistant for aquaculture). As the principal engineer of Mas Ahya, I maintained and improved the core system, releasing it for internal farmer use. It was presented at Microsoft AI Day by Microsoft CEO Satya Nadella and featured in various news outlets and social media platforms

• Developed internal LLM application, enabling the internal shrimp team to receive better recommendations on which ponds are ready for harvest via conversational interactions.

• Initiated and led the adoption of LLM-based automated testing and evaluation to strengthen engineering processes, particularly in QA testing and performance assessment of LLM applications.

• Inspired the development of an OCR LLM-based solution that reduced and saved engineering costs.

• Planned LLM Apps MultiAgent systems roadmap and conducted proof of concept (PoC) with capabilities integration with WhatsApp chat non-text media, including voice (text to speech, speech to text), image, and multi-step data confirmation.

• MLOps

• Achieved up to 80% cost reduction in LLM services operations through performance optimization and resource efficiency improvements.

• Led cloud cost analysis, implementing strategies that resulted in up to 60% monthly cost savings through usage optimization.

• Ensured smooth deployment and operational efficiency of AI services.

• Conducted AI service assessments, identifying and addressing areas for cost reduction and performance enhancement.

• Contributed to designing end-to-end Edge Devices and AI roadmap for seamless integration and deployment.

• AI Squad Leader

• Squad & Tech Implementation Leadership

• Led ML/AI development driving cross-functional collaboration with engineers, PMs, and stakeholders.

• Managed sprint planning, reviews, and retrospectives to ensure efficient Agile execution.

• Oversaw the implementation and maintenance of ML/AI services across AWS, GCP, Azure.

• Enforced code review standards and integrated automated quality control tools for reliability and scalability.

• Ensured seamless ML service deployment across cloud environments.

• Technical Council Contribution

• Regular speaker to the Tech Council, a bi-weekly multi-division forum involving VP, engineers, and PMs.

• Critiqued and reviewed technical proposals and implementation strategies to ensure optimal execution.

• Supervised AI Projects

• Optical Character Recognition Solutions

• Automated sensitive data masking for bank statements (BNI, BRI, BCA, CIMB), processing up to 500-page documents with RabbitMQ for efficient workload distribution and manual error reduction.

• OCR enhancement with GPT-4.0, cutting processing time, manual data collection, and engineering costs.

• Document detection and classification for eKYC, utilizing YOLOv8 to optimize ID verification while reducing unnecessary OCR processing costs through manual filtering.

• Custom OCR system with Azure Document Intelligence, replacing third-party services and cutting costs.

Tools used: Python, Langchain, Langgraph, Langfuse, FastAPI, AWS, GCP, Azure Platform, Nomad, Kubernetes, Grafana, GPT, LLM, Ollama, Huggingface, Alembic, Pandas, Polar, PyTorch, PostgreSQL, MongoDB etc.

• OCR Based Solution

• Developed OCR-based desktop applications to enhance the efficiency of banana harvest form record-keeping using GPT-4o, ensuring easy maintenance by optimizing prompt engineering assets.

• Developed API-based OCR Services leveraging GPT-4o for seamless integration and communication with other services.

• LLM Based Application

• Built and Deployed End-to-End RAG-based (Retrieval Augmented Generation) applications using GPT-4o with a Streamlit web application to provide internal educational resources on banana cultivation, backed by grounded data knowledge.

Tools used: Python, FastAPI, Langchain, OpenAI, Streamlit, Pandas, PyQt5, Docker

Role: Machine Learning & Software Engineer, DevOps. Joined as a Machine Learning Engineer and then my role grows into Software Engineering in general and Development Operations (DevOps)

• Machine Learning related responsibility

• Developed a user clustering service for better analyzing data.

• Developed product recommendation service based on similar product and caching strategy for its output.

• Developed image understanding related functionality such as object detection for duplicate image search, object detection for style card generator functionality (outfit reference image consists of products to be recommended), color picker based on similar RGB value etc.

• Developed service to search for product by its similar characteristic based on image and/or tagging, used for other functionality such as user wardrobe (user’s personal collection and recommended outfit), stylecard generator (admin tool for faster generation of stylecard) and product similar for main website shop page.

• Developed collaborative filtering to recommend products based on user feedback and construct generic user profiles characteristic.

• Software Engineer related responsibility

• Built and maintained machine learning, image and data API services using Python Fast API.

• Built and maintained task scheduler services using Celery, Django, Mongo db and Rabbit MQ

• Responsible for building and maintaining Python API services and contributing to Node JS API maintenance.

• Worked on the creation, design and maintenance of asset management services with image manipulation capabilities (padding, cropping, resizing, changing color) and caching it’s derivative image for faster load.

• Designed, maintained and implemented a message broker for communication between services using Rabbit MQ.

• Contribute to the debugging and migration of the Node JS API application to the new framework.

• Adding listener and notifier Postgres database for data validation.

• Responsible for migrating data, cleansing data to database or serving as a data summary for further analysis.

• DevOps related responsibility

• Manage and handle all things related to development operations, such as deployment, server management, application containerization, automation, etc.

• Initiate and implement the use of Kubernetes for application deployment along with CI/CD pipeline deployment using Jenkins, Ansible and other related frameworks.

• Dockerize and containerize all Yuna applications for microservices application deployment into GCP and AWS.

• Worked on creating, designing, testing and maintaining infrastructure for all services and microservices-based applications using Kubernetes.

• Handle network traffic using Ingress, Nginx, Load balancer and Cloudflare.

• Built automation task such as selenium testing, Telegram Bot - Jenkins integration (able to send Jenkins command via button Telegram chat group updating images product set on Google Vision, automatic bakcup database, WhatsApp server for sending message via UI automation (web based and windows apps based

Tools used: Python, Scikit-learn, OpenCV, Numpy, Pandas, Flask, FastAPI, Django, Celery, Rabbit MQ, Postgres, Mongo DB, NodeJS, TypeORM, Typescript, Google Cloud Platform, Amazon Web Services, Google Vision, Kubernetes, Ansible, Terraform, Kubernetes Operations (KOps), Jenkins, Cloudflare, Telegram bot UI, Bitbucket webhook, Selenium, PyWinAuto, Bash Script.

Intelligent Traffic Monitoring System is an image processing based application for vehicle counting (cars, buses, trucks, motorcycles, pedestrians) and traffic violation detection, including illegal turns and wrong-way movement. The system processes video input from online streams, video files, or live cameras to perform real-time analysis. This project was developed by a team of two, with the following contributions on my part

• Became the leader for project completion.

• Provided technical guidance and practical implementation ideas, including suggested methods, tools, and development workflows.

• Contributed to dataset creation using low resolution CCTV images for training purposes.

• Prepared, tuned, and trained the object detection model (YOLOv5) using the constructed dataset.

• Developed a web-based interface for the system using the Flask framework.

• Implemented the front-end interface using HTML, CSS, and JavaScript.

• Built the traffic violation detection algorithm by combining video processing techniques with object detection outputs.

Tools Used: Python, Tensorflow, PyTorch, OpenCV, Flask, Javascript, HTML/CSS

This project was my undergraduate final project at Universitas Gadjah Mada, majoring in Electronics and Instrumentation (Department of Computer Science and Electronics). The goal was to develop a method for recognizing Quranic reading laws from verse images and to build a web-based interface to display each step of the processing pipeline. This project consists of the following work

• Built a dataset of Arabic characters from cropped images of different variations of Arabic letters in 10 types of fonts. Implemented by creating a simple tool for automatic cropping and saving with a user interface created using the PySimpleGUI library in Python.

• Applied image processing techniques to extract the characteristics of Arabic letters in the image of Al-Quran verse needed for the purposes of determining the reading law, including character detection that marks the occurrence of the reading law by using template matching, designing algorithms for character segmentation using image pixel values, and performing Arabic character recognition using Convolutional Neural Network (CNN) from the created dataset.

• Developed a web-based UI using the Flask framework to integrate reading law recognition methods built in Python with a user interface built in html, css and javascript.

Tools Used: Python, Tensorflow, OpenCV, Flask, PySimpleGUI, Numpy, Pandas, Javascript, HTML/CSS

• Developed machine learning models for detecting and classifying road damage using image datasets.

• Targeted two output types, Pixel-level segmentation producing black-and-white masks highlighting damaged regions and Object detection producing bounding boxes and class labels for each damage type.

• Evaluated and compared multiple state-of-the-art segmentation architectures, including LinkNet, UNet, and FPN, with varied backbones ResNet34, InceptionResNetV2, DenseNet201, EfficientNetB7, and SEResNeXt50.

• Implemented object detection using YOLOv4 and YOLOv5 for classification and localization of road damage.

• Conducted the entire research, experimentation, and model comparison independently within a one-month timeline.

Tools Used: Python, Tensorflow, PyTorch, OpenCV, Google Colaboratory

The project was part of the Student Creativity Program (PKM) funded by the Directorate General of Higher Education (DIKTI) Indonesia. We designed a smart barrier system for pedestrian pathways, using real-time object detection powered by a Raspberry Pi 3 and Intel Movidius Neural Compute Stick 2. The system automatically identifies pedestrians and cyclists while blocking unauthorized vehicles from entering the area. This project was carried out by three people and the work I did

• Served as the team leader, coordinating project development and deliverables.

• Designed the hardware concept, component layout, and integration between the barrier mechanism and the vision system.

• Developed the software pipeline using Python, OpenCV, and Intel Movidius NCSDK.

• Designed detection logic and tuned a pre-trained Convolutional Neural Network model for accurate real-time object classification.

Tools Used: Python, Tensorflow, OpenCV, Raspberry Pi, Intel Movidius, NCSDK

This project was carried out during on the job training at Meteorological, Climatological, and Geophysical Agency (BMKG). Main objective of this project is to make a low-cost environmental monitoring device with various features and combine it with a tool that collects rainwater automatically. The automatic rainwater collector has a cover for the rainwater collection container which only open when it rains. Environmental condition data collected are air humidity, ambient temperature and dust particles. This project was done by three people and the work I did

• Became a leader in technical project completion

• Built a wireless data logger for data collection and communication tools.

• Worked on designing and creating a web-based user interface (html, css, ajax, javascript).

• Worked on designing and making data communication between the device created with the database on a local wireless computer using wifi (C++, mysql, PHP).

• Worked on designing the enclosure of the tools using acrylic (CorelDraw).

• Collaborated on hardware programming (ESP32).

Tools Used: C++, ESP32, PHP, MySQL, Javascript, Ajax, HTML/CSS, Corel Draw

Relink Automatic Plastic Bottle Separator is a vision-based sorting system that uses a Raspberry Pi and camera to classify plastic bottles by color. The system controls a conveyor belt and pneumatic cylinder to automatically redirect each bottle into the correct category. Based on the color detected through computer vision, the Raspberry Pi triggers the actuator to position each bottle into its designated output channel.

My contributions included writing the core Python/OpenCV code, configuring the Raspberry Pi as the main controller, and performing end-to-end system testing to ensure precise and consistent bottle sorting performance.

This project was developed while I was part of the Elins UGM department’s internal organization (HMEI), aimed at helping members easily access the cleaning picket schedule and receive automated reminders through the group’s Line chat.

• This was my first “hello world” project that was actually used by other people, which sparked my interest in building useful tools and automations.

Created with the help of Line Messaging API with a server deployed on a free web hosting (PHP).