Maulana Firdaus
Menu

Mohamad Maulana Firdaus Ramadhan

AI Engineer

I build Voice AI, agentic RAG systems, and machine-learning applications. I work on model integration, backends, and deployment on AWS.

View projectsDownload CV
Mohamad Maulana Firdaus Ramadhan
  • Voice AI
  • Agentic AI
  • Machine Learning
  • Data Engineering
  • Computer Vision

About

I'm Mohamad Maulana Firdaus Ramadhan, an AI Engineer with experience at Neuram and AI Center ITB. My academic background is in Information Systems and Technology at Institut Teknologi Bandung.

My work covers LLM orchestration with LangChain and LangGraph, hybrid retrieval with Qdrant, AI output evaluation, and asynchronous backends on AWS. I also work on computer vision and signal classification on edge devices.

I led Data Community HMIF ITB. I share data science through workshops and speaking sessions. I also compete in Hology, GEMASTIK, and FindIT with teams.

Technical Skills

  • Python
  • PyTorch
  • Hugging Face
  • LangChain
  • FastAPI
  • Docker
  • PostgreSQL
  • Redis
  • Qdrant
  • NumPy
  • Pandas
  • TypeScript
Languages
Python, SQL, C++, Shell Scripting, JavaScript, TypeScript
AI & GenAI
LLMs, Agentic RAG, Hybrid Retrieval (Dense + Sparse, BM25), Text-to-SQL, Transformer Architecture, LoRA, QLoRA, GRPO, Quantization (GPTQ), Prompt Tuning (GEPA)
AI Frameworks
PyTorch, Hugging Face Transformers, LangChain, LangGraph, Scikit-learn, TRL
Backend & APIs
FastAPI, REST APIs, Webhooks, SIP, Twilio, ElevenLabs, Asynchronous Processing, Event-Driven Systems
Cloud & MLOps
AWS (EC2, S3, ECS, SQS, EventBridge, Secrets Manager, CodePipeline), Kubernetes (K8s), Docker, CI/CD, GitHub Actions, MLflow, Airflow
Databases
PostgreSQL, Redis (Queues), FAISS, Chroma, Pinecone, Qdrant, Drizzle ORM
Data & Vision
Pandas, NumPy, OpenCV, YOLOv8, Whisper, K-Means, TF-IDF, Tableau
AI Evaluation
Langfuse, LLM-as-a-Judge, Blind Human Review, BLEU, ROUGE-L, METEOR, BERTScore

Selected Projects

ML systems · Local platform

FlightOps Sentinel

An end-to-end ML system predicting departure delays over 15 minutes at T−24 hours. The pipeline uses a PostgreSQL feature mart, chronological evaluation, MLflow, Airflow, and a FastAPI prediction service.

Role & contribution
Developed the data pipeline and prediction service with feature cutoffs to prevent leakage, experiment lineage, model approval, and Docker-based local deployment.
Results & context
On a 2,206-flight holdout, logistic regression scored ROC-AUC 0.4836 and historical route delay rate scored 0.5437. The platform runs locally with experiment tracking, lineage, and model approval.

ML architecture flow

  1. PostgreSQL
  2. T−24h features
  3. Training & evaluation
  4. MLflow registry
  5. Batch / FastAPI

From data processing to batch and API predictions.

  • Python
  • PostgreSQL
  • MLflow
  • Airflow
  • FastAPI
  • Docker
View codeView documentation

Voice & LLM · Team project

BattleTalk

An English-speaking practice platform with 1v1 battles, gamification, and AI feedback. Its Python service uses Whisper for transcription and Qwen2.5 through Ollama for conversational feedback and scoring.

Role & contribution
Built BattleTalk with the AIce Cream team for the FindIT UGM Hackathon. The application combines speaking practice, speech transcription, and LLM feedback.
Results & context
The AIce Cream team reached the Top 10 at the FindIT UGM Hackathon 2025 with BattleTalk.
  • React
  • FastAPI
  • Whisper
  • Qwen2.5
  • Supabase
View codeView documentation
Original BattleTalk banner showing its AI speaking-practice interface

Applied AI & IoT · Team project

Agrotech

A farm-monitoring PWA combining field-sensor data, image-based plant-disease classification, and AI consultation. Its FastAPI backend includes PyTorch inference and endpoints for sensor and commodity data.

Role & contribution
Collaborated in a four-member team to combine field monitoring, plant diagnosis, and consultation in a web app.
Results & context
An AI and IoT prototype connecting plant-disease classification, field-sensor data, and consultation in one app.
  • Next.js
  • FastAPI
  • PyTorch
  • IoT
View code
Original Agrotech banner showing field-monitoring and plant-disease diagnosis screens

AI product engineering · Personal project

Finance Tracker

A personal-finance app for accounts, transactions, budgets, and financial goals, with an AI assistant and receipt extraction into transaction drafts.

Role & contribution
Built a Next.js app with PostgreSQL and Drizzle, plus LangChain and LangGraph orchestration through Ollama. The assistant reads user data, prepares drafts, and saves transactions after user confirmation.
Results & context
Saves conversation history, validates data ownership, and confirms transactions atomically. The screenshot shows demo accounts and transactions.
  • Next.js
  • TypeScript
  • LangGraph
  • Ollama
  • PostgreSQL
  • Supabase
View codeAI assistant workflow
Finance Tracker dashboard displaying financial summaries from demo data

LLM application · Team project

ChatDoctor

An Indonesian-language health-information chatbot prototype with a chat interface and Python inference service.

Role & contribution
Built ChatDoctor with the team for the Data Science National Tournament, using Sailor2-1B-Chat through Hugging Face Transformers.
Results & context
A Top 10 Finalist at the Data Science National Tournament 2024 with an LLM health-information conversation prototype.
  • Next.js
  • Python
  • Hugging Face
  • Sailor2
View code

Computer vision · Team prototype

RoadEye

A traffic-monitoring prototype combining CCTV views, a map, and accident reports. Its design includes anomaly detection and traffic-density analysis.

Role & contribution
Built RoadEye with the M2VR team for GDSC Hackfest Indonesia. The app combines a Next.js interface with a Python and FastAPI backend.
Results & context
A Top 20 team at GDSC Hackfest Indonesia 2024 with a traffic-monitoring prototype.
  • Next.js
  • FastAPI
  • Python
  • PostgreSQL
View code
Original RoadEye screenshot showing CCTV monitoring and a traffic map

Data mining · Team project

Marketplace Recommender System

A TokoGenZ solution that groups products, recommends similar items, and helps sellers explore categories, prices, sales, and ratings through a Tableau dashboard.

Role & contribution
Worked with the SySendiri team on product data using a lexicon and TF-IDF, K-Means clustering, and NearestNeighbors recommendations based on product-name similarity.
Results & context
1st place at Hology 7.0 Data Mining. The model groups products into 11 clusters and recommends items by content similarity.
  • Python
  • K-Means
  • TF-IDF
  • NearestNeighbors
  • Tableau
View presentation
Original TokoGenZ presentation slide containing the seller analytics dashboard and chart explanations

Computer vision · Team project

Trash Reporting System

A Garbage Report prototype for detecting trash in uploaded images, extracting locations, and preparing reports for sanitation staff.

Role & contribution
Developed the solution with a team for GEMASTIK XVII, including custom YOLOv8 training for trash detection.
Results & context
A national Top 20 Finalist at GEMASTIK XVII 2024. YOLOv8 evaluation scored mAP50 0.92 on a dataset of 3,669 training, 260 validation, and 258 test images.
  • Python
  • YOLOv8
  • Computer vision
Read paperFinalist certificate

LLM post-training · Experiment

GRPO Indonesian Legal Reasoning

A post-training experiment for Indonesian legal reasoning using Qwen2.5-14B-Instruct, LoRA, and Group Relative Policy Optimization.

Role & contribution
Prepared the GRPO training notebook, reward function, dataset splits, and reasoning-data generation script.
Results & context
Evaluates model outputs with BLEU, ROUGE-L, METEOR, and BERTScore, and analyzes hallucinations and reasoning completeness on IndoLaw documents.
  • Qwen2.5
  • LoRA
  • TRL
  • GRPO
  • Python
View codeTraining notebookRead paper

Experience

  • Apr 2026 — Present

    Bandung, Indonesia

    AI Research Assistant

    AI Center ITB

    • Built an agentic AI platform using LangChain and LangGraph for multi-step LLM orchestration.
    • Developed a hybrid RAG pipeline with Qdrant, dense embeddings, BM25 retrieval, and citation-grounding controls.
    • Implemented prompt versioning and an offline GEPA-based prompt-tuning workflow to evaluate and optimize complete LLM prompt templates on frozen evaluation cases.
    • Integrated local and hosted LLM providers, including Ollama and OpenAI-compatible APIs, through configurable Python services.
    • Added Langfuse tracing, reproducibility metadata, and persisted artifacts for debugging, monitoring, and experiment audits.
    • Built an LLM-as-a-Judge and blind human-review evaluation workflow using frozen, pseudonymized evaluation packets to assess output faithfulness, actionability, coherence, completeness, and inter-rater agreement.

  • Feb 2026 — Present

    Jakarta, Indonesia

    AI Software Engineer

    Neuram

    • Scaled production Voice AI systems to support over 40,000 conversations and 1,500+ hours of active voice traffic by integrating telephony, AI agents, and event-driven backend workflows. The appointment workflow integrates SIP INVITE protocols from client domains with Neuram's Twilio infrastructure and ElevenLabs webhooks.
    • Reduced perceived conversational latency by 60% by optimizing real-time audio streaming and response handling for faster initial playback.
    • Reduced perceived inbound pickup latency by over 80% by parallelizing call initialization and backend provisioning while triggering immediate pre-opening audio.
    • Reduced database response latency by 96% and API payload sizes by over 96% by optimizing SQL queries, data parsing, and payload structures.
    • Integrated external APIs to retrieve customer profiles and available slots, then injected them as dynamic variables into ElevenLabs agents for personalized real-time conversations.
    • Eliminated manual post-call reporting delays by automating conversation processing, structured analytics generation, and delivery of results to client systems. Post-call API callbacks deliver call outcomes and data insights to client systems immediately after session termination.
    • Implemented ElevenLabs server-side tools to execute business logic and interact with external systems during live voice conversations.
    • Reduced LLM operational costs by over 88% while maintaining production-level performance by optimizing model deployment and serving architecture across staging and production environments.

  • Oct — Dec 2025

    Bandung, Indonesia

    AI Research Assistant (RAISA Project – SKK Migas)

    AI Center ITB

    • Increased manual data-query accuracy from 28–33% to 82.7% by implementing a deterministic response layer using Qdrant semantic retrieval and structured query templates. The retrieval implementation also uses Qdrant Hybrid Search with dense and sparse approaches.
    • Increased out-of-scope rejection precision from 88.3% to 99.0% across 103 negative test cases by implementing strict-match intent validation and retrieval safeguards. Additional model evaluation used an automated LLM-as-a-Judge workflow across 1,000+ samples.
    • Improved recurring-query reliability by developing deterministic SQL template routing that bypassed generative execution and directly bound user variables to validated query templates. The Text-to-SQL workflow is orchestrated with LangGraph.
    • Strengthened SQL output reliability by integrating an independent Evaluator Agent to validate generated results against user intent before response delivery. The enterprise Agentic RAG system supports complex QnA and high-precision data extraction from unstructured Oil & Gas documentation.
    • Built an asynchronous Excel ingestion pipeline for batch uploads, overrides, and updates to template question-response pairs in the knowledge base.

  • Jul — Sep 2025

    Jakarta, Indonesia

    AI Engineer Intern

    PT. Solutopia Untuk Nusantara (SPUN)

    • Automated 85% of repetitive customer inquiries by architecting and deploying a production AI assistant for global visa services. Model optimization supports efficient inference.
    • Reduced customer response time from a 30-minute manual baseline to under 5 seconds by automating multi-turn customer-support interactions.
    • Supported a 10x surge in concurrent requests by designing an asynchronous AWS architecture using ECS, SQS, and EventBridge. The infrastructure also uses EC2, Amazon S3 for document storage, and Secrets Manager for credential management.
    • Improved response grounding for dynamic visa regulations by engineering a multi-source retrieval pipeline using Pinecone and PostgreSQL.
    • Reduced deployment cycles to minutes by automating build and release workflows through AWS CodePipeline.

  • Mar — Dec 2025

    Bandung, Indonesia

    AI Research Assistant

    Bandung Institute of Technology

    • Achieved 100% defect detection accuracy for partial-discharge signals by developing a real-time LSTM-Attention classification system with faculty researchers. The architecture also uses Dense Concatenation.
    • Reduced local inference latency to milliseconds by optimizing and deploying signal-processing models on FPGA and Raspberry Pi 5 for real-time industrial monitoring.
    • Built a 2,312-sample partial-discharge signal dataset by engineering the acquisition, cleaning, and preprocessing pipeline using Analog Discovery 2.
    • Enabled real-time equipment-health monitoring by deploying an end-to-end dashboard for predictive status and signal visualization.

Organizational Experience

  • Jun 2024 — May 2025

    Bandung, Indonesia

    Head of Community

    Data Community HMIF ITB

    • Led data science and machine-learning activities for the student community. Coordinated technical workshops and events to develop members' data engineering and analytics skills.

  • Sep 2023 — Aug 2024

    Bandung, Indonesia

    Curriculum Developer

    Google Developer Student Club ITB

    • Designed a Machine Learning curriculum, roadmaps, and hands-on tasks for 800+ members. Created project-based assignments so members could apply the material.

  • Nov 2022 — Oct 2023

    Bandung, Indonesia

    Database Division Staff

    STEI-K Student Council

    • Managed student data and produced analytics reports using Python, Pandas, and Excel. Received Best Staff of the Month in February 2023.

Speaking Engagements

2024

Symphony of Data: The Key to Brilliant Decision Making

KAT ITB · Guest speaker

  • Presented how data analytics supports decision-making at a centralized training event.
View certificate

2024

Data Competition Strategies

Kaderisasi Wilayah STEI-K ITB 2023 · Guest speaker

  • Shared data-analysis and machine-learning strategies for student competitions.
View certificate

2024

Introduction to Data Science

SPARTA HMIF 2023 · Guest speaker

  • Delivered an introductory session on data science concepts and career paths for new students.
View certificate

Papers & Research

2025

Research manuscript · Speech & LLM

Deteksi Penipuan Suara Berbahasa Indonesia: Studi Komparatif Antara Arsitektur LALM dan ASR+LLM

Read document

M. Maulana Firdaus R., Atqiya Haydar Luqman, M. Fauzan Azhim

Compares Qwen2-Audio-7B with a Whisper and Sailor2-8B pipeline for detecting fraud in Indonesian conversations, using QLoRA and synthetic audio.

Evaluation
On 104 synthetic test samples, ASR+LLM reached F1 99.03% and LALM 95.19%. Real-audio analysis identified a baseline transcription error.

2025

Research manuscript · LLM post-training

Penyusunan Amar Putusan Berbasis Reinforcement Learning untuk Meningkatkan Reasoning Hukum Pengadilan Indonesia

Read document

Mohamad Maulana Firdaus Ramadhan, Rizqika Mulia Pratama, Ikhwan Al Hakim, Rinaldi Munir

Experiments with Qwen2.5-14B-Instruct, LoRA, and GRPO on IndoLaw documents. The reward combines textual similarity, format compliance, and reasoning completeness.

Evaluation
Measures BLEU, ROUGE-L, METEOR, and BERTScore, then analyzes hallucinations and reasoning completeness in model outputs.

2025

Research manuscript · Legal NLP

Generasi Amar Putusan Hukum Indonesia dari Dokumen Pengadilan Menggunakan Large Language Model

Read document

Mohamad Maulana Firdaus Ramadhan, Rizqika Mulia Pratama, Ikhwan Al Hakim, Ayu Purwarianti

Generates draft legal verdicts from five sections of Indonesian court documents using Sailor2-1B and QLoRA under limited computing resources.

Evaluation
On 50 test documents, the model scored ROUGE-L 0.6267 and BERTScore F1 0.8212 for verdict-text similarity.

2025

Research manuscript · Computer vision

Animal Species Classification Using Deep Learning

Read document

Rizqi Andhika Pratama, Khayla Belva Annandira, Mohamad Maulana Firdaus R.

Compares VGG16, ResNet50, EfficientNet-B3, ViT-B16, and YOLOv11-cls on Animals10, with image-classification integration into EyeNimal.

Evaluation
On Animals10 validation with an 80:20 data split, EfficientNet-B3 reached 98.38% accuracy and YOLOv11-cls 97.84%.

2025

Research manuscript · Applied machine learning

Penerapan Pembelajaran Mesin untuk Prediksi Keputusan Investasi pada UMKM

Read document

Arvyno Pranata Limahardja, David Dewanto, Ricky Wijaya, Nicolaas Heru Dreandachrista, Mohamad Maulana Firdaus Ramadhan

A Shark Tank US study combining financial and demographic features with business-description embeddings. Compares classifiers, Optuna tuning, and SHAP interpretation for deal decisions.

Evaluation
LightGBM reached about 74% accuracy on 90 test samples from the Shark Tank US dataset.

2024

Research manuscript · Computer vision

Pelaporan dan Pendeteksian Adanya Sampah Guna Menjaga Kebersihan Lingkungan Berbasis YOLOv8

Read document

Mohamad Maulana Firdaus Ramadhan, Rizqika Mulia Pratama, Ikhwan Al Hakim, Ayu Purwarianti

Computer vision research on YOLOv8 trash detection in images, location extraction, and a reporting workflow through Garbage Report.

Evaluation
YOLOv8 reached mAP50 0.92 after 100 epochs, with 3,669 training, 260 validation, and 258 test images. Error analysis identified false positives on trees and rocks and missed small trash objects.

2023

Research manuscript · NLP

Pemrosesan Bahasa Alami Untuk Analisis Sentimen Dengan Menggunakan Bidirectional GRU

Read document

Mohamad Maulana Firdaus Ramadhan, Ikhwan Al Hakim, Rizqi Andhika Pratama, Rizqika Mulia Pratama

Models emotions in Indonesian Twitter text using preprocessing, FastText embeddings, and a Bidirectional GRU across five emotion classes.

Evaluation
Evaluation of five-class emotion classification on Indonesian Twitter text scored accuracy 0.70, average precision 0.72, and average recall 0.71.

2023

Research manuscript · Computer vision

Monitoring Otomatis Kondisi Lalu Lintas untuk Mendeteksi Adanya Kecelakaan dengan Menggunakan YOLOv8

Read document

Mohamad Maulana Firdaus Ramadhan, Rizqika Mulia Pratama, Ikhwan Al Hakim, Ayu Purwarianti

Research on detecting accidents in CCTV footage using YOLOv8, with a proposed automatic reporting workflow for detected traffic anomalies.

Evaluation
YOLOv8 validation scored mAP50 0.924430 on 2,220 images, with 10,989 images used for training.

Awards

Certifications

Moments

On stage with award recipients, holding the first-place Data Mining placard and a trophy
20241st place in Data Mining · Hology 7.0
Two award recipients holding a placard and trophy in front of the Hology 7.0 backdrop
2024Award ceremony moment · Hology 7.0
A team presenting its competition work to judges in GEMASTIK XVII documentation
2024Finalist presentation · GEMASTIK XVII
A team presenting feature engineering for network traffic classification
2024Hustler team presentation · Objective Quest, Airnology 3.0
A classroom speaking session with slides about data competition strategies
2024Sharing data competition strategies · STEI-K regional training
Posing with an organizer while holding an appreciation certificate for a data mining competition speaker
2024Speaker appreciation · STEI-K regional training
Three MIQ team members from Institut Teknologi Bandung working together with computers and laptops at a competition desk
At the competition desk · MIQ team, ITB
Best Staff Award documentation for Mohamad Maulana Firdaus in IT Database, with a portrait at the ITB campus
Best Staff Award · IT Database, February

Contact

Get in touch.

Get in touch to discuss AI engineering, ML systems, research, or collaboration. You can also explore my code and experiments on GitHub.