Amrina

Educational technologist and AI developer, building AI that helps people learn and that people can trust.

I hold an M.Sc. in Educational Technology from Saarland University, Germany, and two bachelor's degrees in Informatics and English Education. I build LLM applications with retrieval-augmented generation1, design digital learning experiences2, and study how AI systems can explain things in a way that fits what each learner already knows3. My recent projects bring AI into everyday work: internal assistants that cite their sources, GDPR-aware automation and the step-by-step adoption of AI in organisations.

Based in Saarbrücken, Germany. Open to roles in EdTech, AI and research.

See my projects LinkedIn GitHub

Projects

AI for learning and research, web apps, websites and learning design. Each one links to a live version, a project page or the code.

ERM Tutor-AI

Master's thesis, Saarland University

An AI tutor for the Empirical Research Methods course. It answers from the course's own lecture materials using retrieval-augmented generation, shows students where each answer comes from, and responds like a tutor with reflection and follow-up questions.

Rated 6.75 / 7 by course instructors, against 6.00 for ChatGPT-4o (p < .001)
Answers about 30% shorter and more consistent
Designed in four interface iterations (design-based research)

Python, GPT-4o, LangChain, ChromaDB, Streamlit, RAG

Live app Project page Code

6.75 / 7
instructor rating, against 6.00 for ChatGPT-4o

Metadata Lens overview: completeness of Dublin Core fields per repository

Metadata Lens: quality and knowledge graph for research metadata

Research data project, 2026, rebuilt every week from live repository data

Harvests real publication metadata from the JuSER (Forschungszentrum Jülich) and KITopen repositories over OAI-PMH, checks it against the OpenAIRE guidelines, maps it to Schema.org and DCAT and builds an RDF knowledge graph that you can query with SPARQL directly in the browser.

Polite OAI-PMH harvester with resumption tokens; 17 quality rules for identifiers, access rights, licences, dates and languages, each with a severity and a score
Semi-automatic curation: proposed fixes such as ISO dates and canonical licence URIs, duplicate detection and publisher spellings, accepted or rejected by a person and exported as JSON
Weekly GitHub Actions pipeline and an Airflow DAG; Next.js and TypeScript app with Oxigraph (WebAssembly) for SPARQL; tests for the pipeline and every example query

Python, OAI-PMH, Dublin Core, Schema.org, DCAT, RDF, SPARQL, Next.js, TypeScript, Airflow

Live app Project page Code

800 records
harvested from two research repositories and rebuilt every week

Range Atlas map: records of the Italian locust over a climate suitability surface across Europe

Range Atlas: species distribution maps from open biodiversity data

Research data project, 2026, rebuilt every week from GBIF and OBIS

Maps where ten European grasshoppers and crickets, and two marine species, have been recorded since 2000 and models where the climate suits them. Every step from the raw records to the map is visible: quality checks, model scores, run history and the data itself.

Darwin Core records from the GBIF and OBIS APIs in a Parquet lakehouse (local or S3); quality checks in DuckDB SQL; WorldClim predictors attached with Dask
GLM and Random Forest against a target-group background, scored with spatial block cross-validation; response curves and predictor importance per species
FastAPI on PostGIS with vector tiles (ST_AsMVT, cached, ETags) and model runs locally or on AWS Batch; React, TypeScript and MapLibre app with DuckDB in the browser; tests in CI, weekly GitHub Actions rebuild

Python, DuckDB, Parquet, Dask, scikit-learn, PostgreSQL/PostGIS, FastAPI, React, TypeScript, MapLibre, Docker, AWS (S3, Batch)

Live app Project page Code

12 species
mapped from GBIF and OBIS records, ten of them modelled

RMN AI Assistent: chat answer with a receipt of the documents and stock records it used

RMN AI Assistent: AI colleague for a restaurant team

AI project, chat assistant and automation, 2026

An AI assistant for the staff of a fictional Indonesian restaurant in Saarbrücken. It answers questions about dishes, allergens, hygiene rules and catering from the restaurant's own documents, looks up stock and free tables through a REST API, proposes supplier orders and turns guest e-mails into checked reply drafts.

RAG with sources: every answer shows which document sections and system records it used, and it refuses when nothing matches
Tool calling with six read-only tools; order quantities, approval limits, opening hours and seat capacity are computed in code, not by the model
Runs with OpenAI or a self-hosted model through an OpenAI-compatible API, or without a model in demo mode; 16 tests, CI and Docker

Python, FastAPI, React, TypeScript, RAG, tool calling, SQLite, Docker

Live app Project page Code

6 tools
read-only; prices, stock rules and capacity are computed in code

Ticket list of SolutionDesk with priority, status and first-response SLA

SolutionDesk: learning platform with a support desk

AI project, incident management, 2026

A digital learning platform and the support desk behind it. Every error carries an error code and a request ID; support analyses incidents with structured logs, read-only SQL diagnostics and a knowledge base, then resolves them or escalates to Hosting or Development.

Ticket workflow with impact × urgency priority, first-response SLA, replies to customers and Jira CSV export
XML course package import with validation and XXE protection; log analysis by error code and request trail
Two planted bugs traced from the customer report to the fix in Java, each with a regression test; 43 tests and CI

Java, Spring Boot, SQL, XML, HTML and CSS, Docker, incident management

Live app Project page Code

43 tests
two bugs traced from the customer report to the fix

Project board of the KI-Einführung platform with status, rank and data protection light

KI-Einführung: AI adoption platform

AI project, AI adoption, 2026

A platform that mirrors the whole work cycle of an AI Solutions Manager: ideas from business units, process mapping, a data protection check, use case assessment, management decisions, testing, a 90-day plan, employee training and an AI radar. German by default, with English as an option.

Five roles (business unit, AI manager, data protection, IT, management) with their own permissions; data protection and IT approve separately before management decides
Personal data replaced before every model call; a knowledge assistant answers only from approved documents and names the source
"AI licence" course with final test, PDF certificate and trainer analytics; decision paper as PDF; 28 tests and CI

Python, Streamlit, LLMs, GDPR, change management, e-learning

Live app Project page Code

5 roles
data protection and IT approve before management decides

Internal AI Ops Assistant

AI engineering and automation, 2026

An internal AI service for a healthcare company: it answers policy questions with sources, turns incoming documents into validated data and triages Jira tickets to Microsoft Teams. Personal data is removed before every model call, and sensitive cases go to a human review queue.

All answerable questions answered with the correct source; 44 of 44 document fields extracted correctly (small self-made test sets)
GDPR pseudonymisation, role-based API keys, HMAC-signed webhooks, audit trail and Prometheus metrics
Connectors for Jira, Teams, Confluence, SharePoint and Nextcloud, plus an n8n workflow; 23 tests and CI

Python, FastAPI, LLMs and RAG, GDPR, Docker, GitHub Actions, n8n

Live app Project page Code

44 of 44
document fields extracted correctly in a small test set

Chart: steering a language model toward the user's perspective

Whose Inference?

Research pilot, natural language processing

The same sentence means different things to different readers. This pilot predicts how a specific reader will interpret a statistical statement from a personal knowledge base of what they know and believe, rewrites the statement so it is not misread, and looks inside a language model to see how it uses that knowledge.

Knowledge-graph retrieval raises reader-prediction accuracy from 42% to 61% (75% with GPT-4o)
RSA-style cooperative rephrasing with a simulated listener
Activation patching and steering on Qwen2.5-0.5B

LLMs, knowledge graphs, Rational Speech Acts, interpretability, PyTorch

Live app Project page Code and report

42% to 61%
reader-prediction accuracy with knowledge-graph retrieval

Home page of the multilingual Fluvenco website built with TYPO3

Multilingual corporate website with TYPO3

Website, TYPO3 CMS, 2026

A website in English, German and French for Fluvenco, a fictional manufacturer of hydraulic components. Besides the site itself, the project covers the work around it: an editor guide, a translation and approval process for launching country websites, SEO and automated quality checks.

Own TYPO3 13 LTS sitepackage: Fluid templates, TypoScript, site sets and two custom content elements
Connected translations, localized URLs, hreflang, sitemap and editor groups per language
Automated QA of all 45 pages: links, SEO tags, alt texts and WCAG 2.1 AA, with 0 issues

TYPO3 13, Fluid, TypoScript, multilingual, SEO, accessibility, Playwright

View website Project page Code and editor guide

0 issues
in automated checks of all 45 pages, including WCAG 2.1 AA

Screenshot of Victoria, the breakfast ordering assistant of Hotel Madeleine

Victoria: breakfast ordering assistant

Web app for a hotel, demo

A chat-style assistant that lets hotel guests order a made-to-order organic breakfast from their phone, and a live dashboard that turns the orders into a printable kitchen sheet. Designed to cut food waste compared with a buffet.

Guests scan a QR code, choose German, English or French, and are guided step by step: room number, breakfast basis, extras and allergies
Orders sync in real time to a PIN-protected staff dashboard
One-click, one-page PDF of all orders for the kitchen

Conversational UX, React, Tailwind CSS, Firebase Firestore, jsPDF, multilingual

Try the demo

3 languages
orders reach the kitchen in real time

How to be SMART about your goals

Instructional design, M.Sc. Educational Technology

A self-paced e-learning unit that teaches students to set SMART goals for their academic and everyday life. I designed the learning objectives, structure and activities, and built the unit as an interactive H5P book.

Clear learning objectives and a short theory section before practice
One section per SMART criterion: self-paced resources, then a task on that criterion
Ends with "Your turn!": learners write a SMART goal for their own life, then a summary

Instructional design, H5P Interactive Book, self-regulated learning, Lumi

Try the learning unit

5 tasks
one for each SMART criterion, then learners write their own goal

Experience

Ten years at the meeting point of education and technology.

2020 to 2025, remote

Website Manager

DKLPT, an association for inter-university collaboration of lecturers, Indonesia

Designed and built the association's website with WordPress, set up domain and hosting, and managed its content for academic audiences for five years.

2023 to 2024

Teaching Assistant, Python

Department of Educational Technology, Saarland University

Supported students in Python courses with exercises and projects, and taught data structures, algorithms and debugging.

2020 to 2022

Web Developer and Marketing Officer

Global Ehsan Relief (NGO), Indonesia

Developed and maintained the organisation's WordPress website: new features, updates, security, backups and performance, and wrote SEO-friendly content so donors could find the organisation online.

2018 to 2020

Educational Mentor and Program Officer

Bank BRI Foundation (YBM), Indonesia

Planned and ran education and aid programmes, and mentored groups of scholarship students.

2016 to 2018

Educational Consultant

Ruangguru, Indonesia's largest EdTech company

Built school partnerships and trained teachers and students to use digital learning tools.

Education

A technical foundation in informatics, and a pedagogical one in teaching and learning.

2022 to 2025

M.Sc. Educational Technology

Saarland University, Germany

1.4 very good, on the German scale where 1.0 is the best grade

Instructional design, multimedia learning, AI in education, empirical methods, eye tracking, machine learning for NLP. DAAD scholarship.

2012 to 2016

Bachelor of Computer Science (S.Kom.)

Syiah Kuala University, Indonesia

3.43 of 4.00

Software engineering, databases, algorithms, web programming.

2012 to 2018

Bachelor of Education (S.Pd.), English

Ar-Raniry State Islamic University, Indonesia

3.4 of 4.0

Pedagogy and curriculum design.

Certificates

Courses beyond my degrees. The full collection is on my Certificate Shelf.

Jul to Aug 2022, 80 hours

Professional Academy: Go programming

Digital Talent Scholarship, Indonesian Ministry of Communication and Informatics (Kominfo) with Progate

Git, HTML and CSS, Go (four modules), a team web project and a final project.

Jul 2022

HTML and CSS course

Progate

Fundamentals of HTML and CSS: four lessons and three exercises. Verify the certificate

Skills

What I use to build, teach and evaluate.

AI and programming: Python, LLMs and prompt engineering, RAG, LangChain, vector databases, PyTorch and Transformers, Streamlit, FastAPI, Docker, HTML and CSS, Go (basics), Git, WordPress, TYPO3, SEO
Learning design and e-learning: Instructional design, design thinking, LMS administration, Articulate 360, H5P, Adobe Captivate
Research and data: Empirical research methods, statistics (ANOVA, regression), eye tracking, Tableau, Google Analytics
Design and collaboration: Figma, Miro, Canva, Illustrator, OBS and Audacity, Jira and Trello, Microsoft 365

Languages

Indonesian native
English fluent
German B2
Arabic basic

Awards

DAAD-ACEH Scholarship of Excellence for master's studies in Germany 2022 to 2024
2nd place, Australia–Indonesia Youth Exchange Program (AIYEP) 2017
Karya Salemba Empat educational scholarship 2015 to 2016

Let's talk

I am looking for roles in educational technology, AI and LLM engineering, learning design, or a funded PhD position. Based in Saarbrücken and available to start right away.

Email me LinkedIn GitHub