Floor 00Home · Lobby

What I bring to a team.

Nine strengths, each backed by shipped work, research or production code. Click any proof to see it in action.

Receptionist“Welcome to Tariq's portfolio! Pick any floor.”

AI engineering

Retrieval that finds the right context

Chunking and retrieval are compared and measured, not guessed.

  • Agentic AI Tutor: 4 chunking strategies stored in separate ChromaDB collections and compared side by side
  • Multi-query expansion (up to 5 queries per question) for better recall
AI engineering

Multi-agent systems

Specialised agents with clear roles, routing and self-checks.

  • LangGraph state machine: intent → router → clarify or personalize → retrieve → tutor → evaluate
  • Tool and protocol work across LangChain, LangGraph and MCP
AI engineering

Measured model quality

I measure LLM systems instead of eyeballing them.

  • 92.4% on the RAGAS framework for the best retrieval configuration
  • ~70% LLM mapping accuracy, measured and published in a peer-reviewed paper
  • 100-question end-to-end eval suite for the digital-twin assistant
AI engineering

LLMs on complex, real-world data

Industrial data is rarely clean. I build pipelines that work reliably with real-world data.

  • Multi-stage OpenAI + DeepSeek pipeline turning raw BACnet data into Industry 4.0 digital twins
  • Tool-calling assistant over live BaSyx twins and InfluxDB, with hybrid retrieval and a numeric-fidelity guard
Product engineering

Production React, TypeScript & Next.js

Fast, well-structured interfaces are where AI work becomes a product.

  • Next.js console with SSR, code splitting and API response caching
  • Kindovio: one React Native / Expo codebase for iOS, Android and web, in 13 languages
  • 4+ years of JavaScript / TypeScript / Next.js across startups and freelance work
Product engineering

AI features people can use

Models become useful when they're in the product, in plain language.

  • AI chatbot giving plain-language access to building data
  • This site's AI twin: a streaming chat grounded in a curated knowledge base
Product engineering

Tested, clean and secure code

Quality is built in from the start, not added at the end.

  • Unit, integration and E2E tests (Jest, Cypress, Detox) across web, mobile and backend
  • Clean architecture with Inversify.js; SOLID and MVC as working principles
  • Kindovio: 212 end-to-end checks; row-level security and column-level privacy for locations
  • Open-source firmware security tooling for the EMBArk analyzer
Delivery & teamwork

Shipping & AI-assisted engineering

From container to live URL, with AI coding tools used deliberately.

  • Docker, CI/CD with GitHub Actions, AWS S3
  • Live, deployed demo: the Agentic AI Tutor on Vercel
  • Daily work with Claude Code, Codex and GitHub Copilot
Delivery & teamwork

Explaining the “why”

Clear writing and demos turn technical work into shared decisions.

  • 4 peer-reviewed publications (VDI Automation, CISBAT, NAMUR)
  • Demoed to Siemens Energy in a 9-person Scrum team
  • English fluent, German zwischen A2–B1
Floor 01About · Studio

From complex data to reliable, production-ready AI products.

I build products end to end: pixel-precise React / Next.js front ends, solid TypeScript and Python back ends, and LLM systems that do real work. That means RAG, multi-agent graphs, and pipelines that transform complex industrial data into digital twins.

Coder“This floor is Tariq's story: M.Sc. done, 3 languages spoken.”

I'm a Full Stack AI Developer at TH Köln and hold an M.Sc. in Digital Sciences (Software Architecture), completed in 2026. At TH Köln I build LLM pipelines that read raw building-automation networks and generate Industry 4.0 digital twins. Before that I shipped web and mobile features at startups in Düsseldorf and Hamburg, and freelanced for clients worldwide.

tariq.config.ts
TM PORTFOLIOSTAFF PASS
Tariq Hussain Magsi
Tariq Hussain MagsiFull Stack AI Developer
Access All floorsClearance LLM ∞
drag me · click to flip
If found, please return to Tariq.

Reward: one well-tested pull request and a very sincere “thank you”.

  • ✓ Allowed on every floor
  • ✓ May feed the AI agents
  • ✗ May not ship without tests
click to flip back

Education

  • 2023 – 2026 · completedM.Sc. Digital Sciences · Software ArchitectureTH KölnThesis: Agentic AI Tutor for Assisting Students in Tackling Complex Assignments
  • 2021 – 2023M.Sc. Artificial Intelligence (partially studied)FAU Erlangen-NürnbergAMOS project: open-source EMBArk firmware downloader, demoed to Siemens Energy
  • 2017 – 2021B.Sc. Computer ScienceDHA Suffa University, Karachi

Languages

EnglishFluent
UrduNative
Germanzwischen A2–B1

Engineering principles

Clean ArchitectureSOLIDOOPMVCDSAClean Code
Floor 02Stack · Engine Room

A full stack that thinks.

How the systems I build fit together, from infrastructure up to the interface. Hover a layer of the 3D model to see the tools I use there.

Engineer“Careful, hot GPUs! LangGraph, FastAPI and Next.js live down here.”

Architecture model · 5 layers request response
Loading 3D model…
Specialty

AI × Full Stack

LangGraph agents and RAG pipelines behind FastAPI, streamed into React and Next.js front ends tested with Jest, Cypress and Detox. One engineer, whole product.

DataEmbeddingsAgentsAPIUI
Reading the model

A request (orange) travels down from the user through the interface, AI, API and data layers. The response (indigo) comes back up. Every layer matches a card below.

13

AI / LLM

Agents, RAG and evaluation

LangGraphLangChainRAGMulti-AgentOpenAI APIDeepSeekChromaDBRAGASPrompt Eng.MCPClaude CodeCodexGitHub Copilot
5

Frontend

Web and mobile interfaces

ReactNext.jsReact NativeExpoThree.js
6

Backend

APIs and services

FastAPINode.jsExpressRESTWebSocketsInversify
9

Data

Databases, streams and IoT

PostgreSQLMongoDBFirestoreSupabasePostGISInfluxDBKafkaMQTTBACnet
5

DevOps

Ship and operate

DockerCI/CDGitHub ActionsAWS S3Git
4

Testing

Unit, integration, E2E

JestCypressDetoxRTL
3

Languages

Daily drivers

TypeScriptPythonJavaScript
Floor 03Journey · Archive

From freelance builds to AI research.

Seven years of shipping software: freelance apps, startup product teams, and now applied LLM research at TH Köln.

Archivist“Seven years of career records. Mind the boxes!”

  • 5roles
  • 2019first client
  • 3German cities
05
Apr 2025 – PresentLLM · Digital Twins

Full Stack AI Developer

Technische Hochschule Köln · Köln

  • Built a multi-stage LLM pipeline (Python, OpenAI, DeepSeek) turning raw heterogeneous BACnet data into homogenized models and auto-generated Industry 4.0 Asset Administration Shells, reaching ~70% mapping accuracy.
  • Applied the generated digital twins to automated performance and energy-efficiency evaluation of building systems.
  • Shipped an AI chatbot on top of the pipeline results for natural-language access to building data.
  • Built the Next.js / React app for the NLP pipeline and BACnet device discovery, using SSR, code splitting and API caching.
  • Co-authored 4 peer-reviewed publications (VDI Automation 2025/2026, CISBAT 2025, NAMUR 2025).
PythonOpenAIDeepSeekNext.jsBACnetAAS
04
Jun 2024 – Mar 2025NLP · IoT

Full Stack AI Developer

Technische Hochschule Köln · Köln

  • Developed a Python backend for storing and processing sensor data (pressure, temperature and 10+ attributes).
  • Designed a digital twin model for a water tank plant.
  • Implemented end-to-end NLP pipelines: ingestion, preprocessing, embeddings and inference.
PythonEmbeddingsInfluxDBMQTT
03
Aug 2023 – Apr 2024Web · Mobile

Working Student · Full Stack Developer

Evolute CX GmbH · Düsseldorf

  • Delivered 5+ features across web and mobile with React, Next.js and React Native, plus 4+ backend features in Node.js and Inversify.js following clean architecture.
  • Wrote unit, integration and E2E tests (Jest, Cypress, Detox) across web, mobile and backend.
ReactNext.jsReact NativeNode.jsInversifyCypress
02
Feb 2022 – Mar 2023Internal tools

Working Student · Software Developer

Pflanzmich GmbH · Hamburg

  • Built the internal dashboard for data and database management, with a React / Next.js front end and a Node.js backend.
ReactNext.jsNode.jsDatabases
01
2019 – 2022Freelance

Full Stack Developer · Freelance

Upwork · Fiverr · Remote

  • Developed and maintained 8+ websites and mobile applications using React, React Native and Expo.
ReactReact NativeExpoFirebase
Floor 04Work · Lab

Projects with real impact.

Ten projects across AI research, product engineering, client work and open source. Pick a project to load its 3D model. Each model shows how that system works: follow the moving data, and drag to rotate.

Scientist“Shh, experiments running. Ten projects, each with a 3D model.”

More on GitHub
3D model · LangGraph agent pipeline1 / 10
Loading 3D model…
Master's thesis · TH Köln · 2026

Agentic AI Tutor

A LangGraph state machine of specialised agents that understands a student's question, adapts to their level, retrieves course material and grades its own answers.

92.4%RAGAS score
Problem

Students stuck on complex assignments get generic answers from chatbots that ignore their level and the actual course material.

What I built
  • Pipeline: understand intent → router → clarify or personalize → multi-query → retrieve → tutor → evaluate
  • Infers competence (novice / intermediate / advanced) live from the student's writing, with no stored profile
  • Four chunking strategies (fixed, recursive, semantic, agentic) in separate ChromaDB collections, compared side by side
  • Ingests YouTube transcripts and web pages, served through a FastAPI backend
Impact

The best retrieval configuration scored 92.4% on the RAGAS evaluation framework.

LangGraphRAGFastAPIChromaDBRAGASPython
Live demo ↗Code ↗
Floor 05Research · Library

Peer-reviewed work on LLM-generated digital twins.

As a developer at TH Köln, I built LLM pipelines that generate digital twins from raw BACnet building networks and then evaluate how those buildings perform. The 3D building at the top of this page is a nod to that work.

Librarian“Four peer-reviewed papers. I've read them all. Twice.”

2026VDI Automation 2026 · Baden-Baden

Evaluating the Mapping Accuracy of Large Language Models for Automated Digital Twin Generation in Building Automation

Proceedings pp. 372–381
2025CISBAT 2025 · Lausanne · J. Phys.: Conf. Series

Automated Performance Evaluation of Buildings Based on a Taxonomy of Building Functions with NLP-Supported Exploration in Building Information Networks

Read via DOI ↗
2025VDI Automation 2025 · Baden-Baden

Automatisierte Performancebewertung der Technischen Gebäudeausrüstung auf Basis einer ganzheitlichen Ordnungsmethode für Gebäudefunktionen

Read via DOI ↗
2025VDI Automation 2025 · Baden-Baden

Automatisierte Erstellung von I4.0-Verwaltungsschalen einer realen Produktionsanlage auf Basis der NE 196

Read via DOI ↗
Floor 06Contact · Rooftop

Let's build something intelligent.

Hiring for full stack, AI engineering, or someone who can do both? Reach out directly, or quiz my AI twin first.

Stargazer“Best view in town. Want to say hello to Tariq?”

LinkedIn GitHub TH Köln