Ninditya Salma Nur Aini — AI Engineer & Data Scientist

Ninditya
Salma Nur Aini

AI Engineer · Data Scientist · QA Engineer

I began with an NLP thesis in 2021, built chatbot flows at Telkom Group, and evaluated LLM quality at Flip. I help close the gap between benchmark scores and systems that stay reliable in real use.

AI Systems Engineering

LLM routing, RAG systems, evaluation workflows, and production-ready architecture decisions.

ML & Data Science

End-to-end ML pipelines with MLflow, Docker, Prometheus, Grafana, and business-driven modeling.

NLP & Bahasa Indonesia

Intent systems, sentiment analysis, entity extraction, and NLP from LSA to BERT.

QA & AI Evaluation

SIT/UAT, Playwright, Selenium, Postman, hallucination checks, and QA-focused model evaluation.

The work. Explained in full.

End-to-end case studies — from problem framing to delivered results.

Featured spotlight

Section 01 · LLM & Generative AI Live

RAG Chatbot Service

"3 design decisions I kept while building a production RAG system." An LLM app with FastAPI, LangGraph, and vector search for grounded knowledge retrieval.

3 Graph nodes

3,072 Embedding dim

4 Response modes

RAG LangGraph Qdrant Gemini

Open case study →

Section 01 · LLM & Generative AI Live

Banking Intent Router

"I reduced banking chatbot LLM cost by 88%." A 4-tier routing system that decides when a query needs a regex filter, a template, RAG, or an actual LLM call — based on confidence from SBERT + LinearSVC.

0.9305 F1 Macro

−88% LLM Cost

~50ms Latency

LLM routing Banking FastAPI SBERT

Open case study →

Section 02 · Applied Data Science Live

Student Dropout Prediction

"1 in 3 students drops out. I built the early-warning system." XGBoost classifier on 4,424 student records — 92.42% accuracy, AUC 97.27%, with Looker Studio dashboard and Streamlit prediction app for counselors.

92.42% Accuracy

97.27% ROC-AUC

32.12% Dropout rate

XGBoost Education SMOTE Streamlit

Open case study →

Section 01 · LLM & Generative AI Live

StudioAI — Generative Image Pipeline

"I built a Stable Diffusion pipeline from first principles." Hyperparameter experiments, CLIPSeg auto-masking for inpainting, multi-step outpainting zoom-out, and a full Streamlit app — Bangkit BFGAI Challenge.

4 Schedulers compared

3 Masking approaches

3GB+ Model managed

Stable Diffusion Diffusers CLIPSeg Streamlit

Open case study →

All projects by domain

Section 01

LLM & Generative AI

Generative AI systems: cost control, prompt optimization, RAG architecture, LLM evaluation, and production-ready applications.

AI stack

Case study Live case study

Banking Intent Router

2025 · Solo project

"I reduced banking chatbot LLM cost by 88%." A 4-tier router that decides when a query is cheap enough for templates, when it needs RAG, and when it must escalate to an LLM.

LLM routing Banking FastAPI

Open case study →

Case study Live case study

HealthTruth - RAG Chatbot Service

2025 · Solo project

"3 design decisions I kept while building a production RAG system." An LLM app with FastAPI, LangGraph, and vector search for knowledge-grounded answers.

RAG LangGraph Qdrant

Open case study →

Case study Live case study

StudioAI — Generative Image Pipeline

2024 · Solo project · Bangkit BFGAI

"I built a Stable Diffusion pipeline from first principles." Systematic hyperparameter experiments, CLIPSeg auto-masking, outpainting zoom-out, and a Streamlit web app — Bangkit BFGAI Challenge.

Stable Diffusion Diffusers CLIPSeg

Open case study →

Case study Live case study

NutriPal AI

2025 · Solo project

"I built a food photo AI that identifies your meal, calculates macros, and warns you about allergens — in one Gemini call." Full-stack nutrition app: Next.js 15, Genkit flows, 4-key rotation, Supabase Realtime.

Next.js 15 Gemini Vision Genkit Supabase

Open case study →

Section 02

Applied Data Science

Data science foundations: time series, MLOps, people analytics, and practical NLP.

Data stack

Case study Live case study

NLP KNN LSA & Chatbot (Thesis 2021)

2021 · Solo project · S.Kom Thesis

"My 2021 thesis. End-to-end from research to product." Classical NLP model (KNN + LSA) integrated as frontend and database of a working chatbot.

Thesis Classical NLP Full-stack

Open case study →

Case study Live case study

Beijing Air Quality Dashboard

2025 · Solo project

"46 MB of sensor data, 5 years, 12 stations — turned into one actionable dashboard." Streamlit app with AQI computation, seasonal trends, meteorological correlations, and geospatial priority mapping.

Streamlit Plotly Chi-square Geospatial

Open case study →

Case study Live case study

Student Dropout Prediction

2025 · Solo project · IDCamp Expert

"1 in 3 students drops out. I built the early-warning system." XGBoost classifier, 92.42% accuracy, AUC 97.27%, with Looker Studio dashboard and Streamlit app for counselor intervention.

XGBoost Education SMOTE Streamlit

Open case study →

Case study Live case study

Employee Attrition Prediction

2025 · Solo project · IDCamp Expert

"Overtime is a 3× risk multiplier. The data proved it." XGBoost HR system with 86.11% recall, Metabase dashboard on Docker + PostgreSQL for operational early-warning.

XGBoost HR Analytics Metabase Docker

Open case study →

Case study Live case study

Sentiment Analysis: ML vs GenAI

2025 · Solo project

"ML beat GenAI on accuracy, speed, and cost — all three." Bidirectional LSTM vs TF-IDF+MLP vs GenAI zero-shot on 11,000 rows of Indonesian Twitter text. 91% vs 78%, $0 vs $82/yr.

LSTM Flask API Bahasa Indonesia NLP

Open case study →

Case study Live on GitHub

Talent Match Intelligence

2025 · Solo project

"I built a 17-CTE matching algorithm in pure SQL." HR analytics system for Paragon — matching candidates against top-performer benchmarks using star schema, PERCENTILE_CONT, conditional aggregation, and a Streamlit dashboard.

SQL PostgreSQL HR Analytics Streamlit

Open case study →

Backend Engineering · KADA Live case study

CineTrack — Cinema Analytics Backend

2026 · Team project · KADA

"I built the SQL analytics layer behind a cinema chain dashboard." Backend API with Fastify (Node.js), PostgreSQL CTEs for revenue, occupancy, and time-slot analytics — plus Redis caching and EXPLAIN ANALYZE query optimization.

SQL Node.js (Fastify) PostgreSQL Redis

Open case study →

Planned case Coming soon

MLOps Earthquake Pipeline

In progress · Solo project

"How I make sure an ML model does not fail silently in production." End-to-end pipeline with monitoring, experiment tracking, and deployment discipline.

MLOps Monitoring MLflow

Coming soon

Case study Live case study

Fine-tune IndoBERT Financial Sentiment

2025 · Solo project

"From LSA to BERT: what changed and what did not." Fine-tuning IndoBERT on Indonesian financial text — 88.4% accuracy vs 72.3% TF-IDF baseline. Domain-specific fine-tuning is the entire story.

IndoBERT HuggingFace Transformers NLTK · spaCy Docker

Open case study →

Section 03

Quality Assurance & Testing

QA is not a side path; it is why my evaluation, reliability, and product thinking stand out. Cards 1–2 describe production QA work at Flip and Telkom — write-ups in progress.

QA stack

Production · Flip Write-up in progress

LLM Chatbot Evaluation Framework

2024–2025 · QA Analyst · Flip

"How do you know your LLM isn't hallucinating on users right now?" At Flip, I built and ran QA for an LLM-based chatbot with RAG — tracking hallucination rate, answer grounding, fallback triggers, and CSAT correlation across releases.

LLM eval RAG QA Hallucination detection Flip

Write-up in progress

Production · Telkom + Portfolio Write-up in progress

AI System Test Automation

2022–2024 · IT Operational · Telkom

"AI systems fail at edges humans don't test for." Playwright + Selenium automation for AI-facing products — multilingual intent stability, regression suite, SIT/UAT workflow for chatbot releases at Telkom Group.

Playwright Selenium pytest SIT/UAT

Write-up in progress

Portfolio project Code complete

ETL Fashion Pipeline

2025 · Solo project

"50 pages scraped, 3 destinations loaded, every step pytest-verified." Modular ETL from BeautifulSoup scraping to 3-destination load: CSV, PostgreSQL, and Google Sheets — with USD→IDR transform, data validation, and full pytest coverage using MagicMock.

ETL PostgreSQL pytest Google Sheets API

Open case study →

Background & skills

I'm Ninditya Salma Nur Aini, based in Yogyakarta, Indonesia. I combine AI engineering, data science, and QA.

I built rule-based chatbot flows at Infomedia (Telkom Group) and later evaluated LLM chatbots at Flip, using RAG-based quality checks for grounding, fallback, and hallucination control.

I trained through Binar Academy and the IDCamp Expert track (Indosat Ooredoo Hutchison), building monitored ML pipelines with MLflow, Docker, and Grafana.

I work from launch to live operations: designing production-ready models and testing them for drift, failure, and real-user reliability.

↗ View CV ↗ Learning Journey Portfolio

Core Stack

AI/ML Core

ML & Data SciencePython, scikit-learn, XGBoost, TensorFlow, pandas, numpy

LLM & RAGLangGraph, Qdrant, FastAPI, prompt optimization, hallucination detection

Computer VisionYOLOv8, object detection, precision/recall/AP@50

NLPIntent classification, sentiment analysis, Bahasa Indonesia, LSA, KNN

Production & Quality

MLOps & DeploymentMLflow, Docker, GitHub Actions, Prometheus, Grafana, Hugging Face Hub, DagsHub

QA & Test AutomationPlaywright, Selenium, Postman, pytest, SIT/UAT, AI stability testing

AI Quality & AuditingKPI dashboards, grounding validation, root-cause analysis, hallucination detection

Infrastructure

Data EngineeringPostgreSQL, MySQL, MongoDB, Redis, REST APIs

Analytics & VisualizationMetabase, Looker Studio, Grafana, Matplotlib, Seaborn

BackendNode.js (Fastify), Python (FastAPI, Flask), JavaScript

NindityaSalma Nur Aini

The work. Explained in full.

Background & skills

Ninditya
Salma Nur Aini