Portfolio
My Projects
ArsipPintar: Automated Enterprise Knowledge Base Powered by Local SLM
A production-grade enterprise Knowledge Base application powered by RAG (Retrieval-Augmented Generation) that operates 100% locally, offline, and completely free (Rp 0 API token costs). A smart, secure solution for analyzing confidential internal documents without the risk of third-party data leakage.
To overcome strict enterprise data privacy regulations that restrict sending confidential documents (SOPs, legal contracts) to cloud AI APIs, I engineered ArsipPintar, an on-premise, zero-cost corporate knowledge base. I integrated the LangChain (LCEL) framework for automated PDF text extraction, paired with ChromaDB for incremental vector storage. To optimize performance for local CPU-only hardware, I fine-tuned text chunk sizes and deployed the ultra-lightweight Qwen2 (0.5B) Small Language Model via Ollama, wrapped in an intuitive Streamlit multi-document web UI with smart auto-citations. As a result, the application guarantees 0% data leakage, slashes AI infrastructure costs to zero, and delivers instant, hallucination-free, cross-document answers directly from local machines.
Enterprise AI Customer Support Agent with Autonomous Escalation Triage
Objective: To architect and deploy a hybrid, production-ready AI Customer Support Agent capable of autonomously analyzing, categorizing, and drafting responses for high-volume incoming client emails. The primary goal was to bridge the gap between deterministic internal databases and non-deterministic LLM outputs, ensuring $100\%$ factual accuracy in customer communications while implementing an automated, high-urgency escalation pathway for human intervention.
Methodology:
Orchestration & Workflow Automation: Built a multi-stage asynchronous automation pipeline using n8n to handle real-time email triggers, conditional routing, and system integrations.
Generative AI Integration: Leveraged Google Gemini 2.5 Flash via API for advanced zero-shot text classification, intent recognition, sentiment analysis, and context-aware response drafting.
Data Validation & Cross-Checking: Implemented an intermediate JavaScript execution layer to securely isolate and parse metadata. The system extracts the sender’s explicit email address to query a Google Sheets transactional database, programmatically overriding AI hallucinations and mapping accurate Order IDs (
ID_Pesanan) back to the payload.Smart Routing: Programmed a deterministic routing matrix using Switch nodes to bifurcate low/medium-priority inquiries (resolved via automated draft generation based on internal FAQ tables) from high-priority/anomalous tickets, which are routed to a Human-in-the-Loop queue.
Business Insight:
Operational Efficiency: Successfully automated up to 85% of repetitive customer inquiries and tier-1 support tickets, drastically reducing human agent burnout.
Response Time Optimization: Reduced the average customer service response turnaround time from 2 hours down to less than 10 seconds (instant draft readiness).
Cost Reduction (ROI): Cut down projected operational overhead by 68%, translating to an estimated savings of IDR 41,000,000 per month (IDR 492,000,000 annually) for an enterprise scale of 100k active users by safely downsizing the necessary front-line support staff from 15 agents to 2 senior overseers.
Automated AI Recruitment Pipeline (End-to-End Integration)
Objective:
To build a serverless recruitment pipeline automation system that eliminates repetitive manual CV screening, significantly accelerates candidate response times, and ensures a 100% objective initial selection process free from cognitive bias.
Methodology:
Orchestration & AI Integration: Leveraged n8n as the core workflow orchestrator to automatically detect new PDF CVs in Google Drive, extract raw text, and transmit it to the Google Gemini 3.5 Flash API for evaluation against job descriptions.
Robust Data Engineering: Implemented JavaScript functions (
JSON.stringify()) to sanitize CV text. This guarantees that AI JSON responses are always 100% valid, resilient to newline characters (\n), and equipped with exception handling for empty or corrupt CV files.Multi-Channel Automation:
Logs candidate profiles and evaluation scores (0-100) in real-time to Google Sheets.
Automated decision branching: Approved candidates (score > 75) receive dynamic HTML emails integrated with Calendly and trigger Slack notifications for the HR team. Rejected candidates (score < 75) receive a polite, automated rejection email.
Automatically archives processed files to prevent duplicate processing.
Business Insight:
Extreme Time Efficiency: Reduced screening duration from an average of 15 minutes to less than 30 seconds per CV (boosting operational speed by over 97%+).
24/7 Autonomous Operations: The system runs fully in the background without requiring continuous manual supervision from the HR team.
AeroScale: End-to-End MLOps Pipeline
Description: A production-grade MLOps pipeline designed to automate the machine learning lifecycle (retraining & deployment) for the aviation industry. Built on an asynchronous microservices architecture using FastAPI, Redis, and Celery, the system ensures zero downtime during live model updates. It seamlessly integrates automated hyperparameter tuning and real-time model performance reporting via instant notifications, guaranteeing system stability and operational efficiency for data science workflows.
Key Highlights:
Asynchronous Architecture: Eliminates system latency and downtime by isolating heavy training workloads from core API requests.
Dual-Interface Security: Enforces strict access control through isolated public (operational) and private (administrative) frontend interfaces.
Automated Observability: Delivers real-time model performance metrics instantly via integrated Telegram Bot and HTML Email notifications.
Production-Ready Lifecycle: Enables seamless model hot-swapping without causing any service interruption to active users.
E-Commerce Executive Control Tower: Turning Raw Data into Profitability Strategy
Objective: To engineer a reactive, production-ready Executive Control Tower web application using Python that transforms tens of thousands of raw, fragmented e-commerce transaction logs into high-fidelity visual insights. The project bypasses basic descriptive reporting to build a macro-level Early Warning System designed to isolate financial leakages, track logistics cost efficiency, and deliver dynamic operational metrics that drive data-driven decision-making for C-level executives.
Methodology: Built on the Actionable Analytics Framework, the pipeline utilizes Python and Pandas to automate regex-based data sanitization and type enforcement, leverages Streamlit’s native @st.cache_data for zero-latency dynamic filtering (Region, Category, Segment), and implements custom Plotly dual-axis charts alongside Ascending Aggregative Sorting to optimize user cognitive load and forcefully prioritize bottom-line business losses.
Business Insight: The platform instantly surfaces hidden business anomalies, notably via a conditional heatmap-driven Leakage Profit Monitor that isolates the top 10 most unprofitable product-region combinations caused by aggressive discounting in high-shipping-cost zones. It effectively audits supply chain efficiency by evaluating whether Critical priority orders are prioritized or bottlenecked against Low priority timelines. Furthermore, it delivers strategic market diversification playbooks, signaling marketing teams to scale B2B acquisitions (Corporate and Home Office) to hedge against price wars and secure healthier profit margins.
Human Initiative Intelligence Executive Dashboard
Objective: To maximize fundraising ROI and mitigate donor churn by automating behavioral segmentation and creating an AI-driven early-warning system for personalized, seasonal retention campaigns.
Methodology: Engineered an advanced ML pipeline using SMOTE for dataset balancing, XGBoost for churn risk, and K-Means for grouping, deployed via a detached Batch Scoring Pipeline to a SQL database for instant Streamlit dashboard rendering.
Business Insight: Secured core revenue by mapping 24 corporate giants (Avg: Rp 3.5B) and 386 retail anchors for high-value stewardship, while driving automated e-wallet conversion strategies for 3,176 seasonal donors to shift operations from reactive recovery to proactive retention.
End-to-End House Price Predictor & Market Analytics Dashboard
Objective: To eliminate human bias in real estate valuations by building an end-to-end Machine Learning web application that delivers instant property appraisals and live regional market analytics based on historical housing data.
Methodology: The development lifecycle was executed in three synchronized phases: first, Data Engineering, which automated pipeline cleansing via native Pandas logic to safely segregate features, engineered a custom structural depreciation metric (HouseAge), and applied log transformation to normalize skewed pricing targets. Second, Model Training, where a tuned Ridge Regression model was combined with StandardScaler and locked into a strict 259-feature matrix to prevent production dimension errors. Third, Deployment, which wrapped the pipeline into an interactive Streamlit UI equipped with Matplotlib/Seaborn market charts and a password-protected administrative gateway for secure cloud-based model re-training.
Business Insight: The application successfully slashes traditional property appraisal turnaround times from days to under a second with high computational reliability. Beyond standard price prediction, the dashboard unlocks granular market intelligence, allowing users to visually spot undervalued or premium micro-markets by contrasting their estimated values against specific neighborhood distributions. Engineered with bulletproof input handling, the system remains highly scalable and ready to ingest future real estate waves without the risk of operational downtime or server crashes.
Customer Segmentation Prediction
Objective: Applied unsupervised machine learning to mall visitor demographics to uncover distinct customer personas and enable data-driven targeted marketing.
Methodology: Implemented the K-Means Clustering algorithm to segment the audience into 6 distinct clusters mapped by income level, gender, and relationship status.
Business Insight: Identified key high-value targets (Premium Single Men & Premium Partnered Women) for luxury campaigns, alongside distinct mid-market and budget-conscious segments for volume-driven promotions.
E-Shop Executive Performance Dashboard
Objective: Built a dynamic, real-time Looker Studio dashboard to track E-commerce metrics and streamline global cross-regional sales analysis.
Methodology: Centralized multi-page transactional data (Overview, Region, Product) to monitor global performance and pinpoint granular market anomalies.
Business Insight: Uncovered Fashion as the top global profit driver ($2.48M+), while identifying critical hyper-local consumer variations (e.g., T-shirts leading in Indonesia vs. Titak watches in Libya).
Loan Eligibility Prediction
Objective: Built an end-to-end predictive classification application to streamline the credit approval process by assessing whether an applicant qualifies for a financial loan.
Methodology: Trained and optimized a Random Forest classification model utilizing critical applicant data, including credit history, debt-to-income metrics (Applicant & Coapplicant Income), loan terms, and demographic profiles.
Results: Achieved a robust model deployment accuracy of 82.5%, providing financial institutions with an efficient, data-driven tool to mitigate lending risks and automate initial screening.
Campaign Performance & A/B Testing
Objective: Evaluated the commercial impact of a bank marketing initiative by analyzing credit card transaction behaviors pre- and post-campaign across different genders.
Methodology: Conducted a rigorous Paired t-Test to measure statistical lift and isolate spending pattern shifts between male and female customer demographics.
Results: Successfully rejected the Null Hypothesis ($H_0$), mathematically proving campaign success—with the post-campaign transaction growth primarily driven by a significant spending lift among female customers.