AI Automation Engineer | Software Engineering Undergraduate | ML Researcher
I am a highly motivated Software Engineering undergraduate with a hands-on builder mentality. I specialize in designing production-grade workflow automations, deploying Agentic AI systems, and implementing parameter-efficient fine-tuning (PEFT) for Large Language Models.
I thrive on bridging the gap between business operational pain points and robust technical solutions—building reliable systems that handle massive data synchronizations, edge cases, and API limits efficiently.
- LLM Infrastructure: QLoRA Fine-tuning, Unsloth, HuggingFace, Model Evaluation (IndoMMLU, ROUGE-L, BERTScore, DIFF-EPITOME)
- Agentic AI & Engineering: Prompt Engineering, Model Context Protocol (MCP), Agent Development Kit (ADK), RAG concepts
- Classical ML & Deep Learning: BiLSTM, Random Forest, K-Means Clustering, TruncatedSVD, CRISP-DM Methodology
- Platforms: n8n (Self-hosted/Cloud), Microsoft Power Automate
- Scripting: Google Apps Script, custom JavaScript code nodes
- APIs: REST API consumption & development, Webhooks, WhatsApp Business API (WAHA)
- Backend: Python (FastAPI, Django REST Framework), Node.js
- Frontend: Next.js 15, React 19, TypeScript, Tailwind CSS
- Infrastructure & DevOps: Docker, Google Cloud Run, Vercel, Linux VPS, Git
- Customer Sentiment Analysis Engine (BiLSTM)
- Engineered an end-to-end NLP pipeline using BiLSTM and FastText Indonesian embeddings to extract insights from >10,000 scraped Google Play Store reviews.
- Tackled minority class imbalance through SMOTE and achieved highly robust test accuracy for real-world text-label contradictions.
- Bank Transaction Clustering & Classification
- Analyzed complex banking data using Unsupervised Learning (K-Means) to identify distinct customer profiles and Supervised Learning (Decision Tree/Random Forest) for cluster prediction with 98% accuracy.
- Automated Document Monitoring System (AIL)
- Engineered a hybrid automation system integrating Google Sheets and Microsoft OneDrive to monitor real-time document pipelines without system timeouts.
- Invented a "Ninja Email Webhook" batching strategy in Power Automate and implemented "Turbo Sync" (Hash Map caching) in JavaScript to synchronize hundreds of rows in under 5 seconds.
- End-to-End WhatsApp Chatbot Workflow
- Deployed an automated data request workflow utilizing n8n and WAHA webhooks.
- Authored custom JavaScript code nodes inside n8n to accurately parse complex JSON message payloads and format dynamic structured bot responses.
- QuickBrief AI – Serverless Text Summarization Agent
- Built a production-ready HTTP REST API via FastAPI, Docker, and Google Cloud Run to summarize large volumes of text in under 2 seconds using Google's Gemini LLM.
- Features robust error handling, edge-case fallback, and scalable zero-config deployment.
- E-Commerce Revenue Insight Agent
- Architected an Agentic workflow integrating the Model Context Protocol (MCP) to securely extract structured data from BigQuery SQL databases.
- Allows the LLM to autonomously retrieve deterministic sales data and generate grounded operational insights.
- CredexAI – Full-Stack AI Credit Scoring Platform
- Developed a frontend-first application for alternative SME credit scoring using Next.js 15, React 19, and TypeScript.
- Integrated Azure AI Language with strict payload validation (Zod) and engineered a deterministic fallback mechanism to gracefully handle AI service timeouts.
- Dual-Stakeholder Recommendation System (TIPDONG)
- Designed a zero-shot cross-domain ML recommendation system (SVD & Random Forest) for a campus food ordering platform, deployed as REST APIs on HuggingFace Spaces.
QLoRA Fine-Tuned Gemma-2 Model for Affective Intervention
International Journal of Advances in Data and Information Systems (IJADIS) – Accepted with Major Revisions
- Authored a research paper on parameter-efficient fine-tuning (PEFT) of the 9-billion parameter Gemma-2 LLM utilizing QLoRA (4-bit NF4 precision) and the Unsloth framework.
- Constructed a synthetic data distillation pipeline formatted into strict Instruction-Input-Output triplets.
- Successfully mitigated catastrophic forgetting mathematically—boosting task-adjacent linguistic stability (ROUGE-L +23.34%, BERTScore +4.79%) while maintaining a robust 60.6% zero-shot accuracy on the IndoMMLU benchmark.