AI Engineer · Cologne, Germany

AI that reasonsover real-worlddata.

I’m Nouman Ahmad. I build LLM and agentic systems, retrieval pipelines and machine-learning models that turn complex operational data into answers and decisions people can trust.

Fig. 0 — My work as a neural network. Hover a node to trace its paths.

0
projects in LLMs, retrieval, agents and machine learning: 0
peer-reviewed papers, one as first author: 0
talks and presentations in Germany, Switzerland and Japan: 0.0
Master’s thesis grade (German scale, 1.0 is best)

§ 01

From automation engineering to AI engineering.

Nouman Ahmad · Cologne, 2026

My work sits where language models meet real operational data: assistants that answer from a system’s own knowledge graph and time series, agents that take over repetitive model-training work, and machine-learning methods that read the condition of equipment from sensor data.

My path started with PLC programming for production machines in Kayseri, continued through mechatronics in Istanbul, and led to AI engineering in Cologne. That engineering foundation shapes how I work: AI systems have to hold up on real, messy data, not only in a demo.

Today I am a Research Associate at TH Köln, where I build LLM assistants and agentic workflows and lead the modKI team’s work on reusable AI models.

Three questions I work on

Education

M.Eng. Automation and ITTH Köln, GummersbachOct 2023 – Jun 2026 · final grade 1.6Thesis (1.3): Development and Evaluation of a Generic Method for Determining the Functional Reserve of Technical Building Components Based on Synthetic and Real Process Data
M.Sc. Mechatronics Engineering (without thesis)Boğaziçi University, IstanbulOct 2021 – Aug 2023 · 3.45 / 4.00
B.Sc. Mechatronics and RoboticsErciyes University, Kayseri2014 – 2019 · 3.47 / 4.00

Languages

EnglishProfessional (C1)
TurkishC1
GermanLimited working proficiency
Urdu & HindiNative

§ 02

What I work with.

The tools and methods behind the projects below, grouped by what they are for.

01 · 9 skills

AI & LLM systems

Agents, retrieval and evaluation for language models.

Agentic AI
LangChain
LangGraph
RAG
Graph-based retrieval
Vector databases
MCP
NLP
LLM evaluation

Data
Retrieval
LLM
Agents
Evaluation

02 · 10 skills

Machine learning

Models for time series, anomalies and prognostics.

PyTorch
TensorFlow
Keras
scikit-learn
Time-series analysis
Anomaly detection
Residual-based prognostics
Synthetic data generation
MLflow
Weights & Biases

03 · 7 skills

Data & analysis

Cleaning, exploring and presenting data.

Python
pandas
NumPy
SciPy
matplotlib
Streamlit
Sensor integrity assessment

04 · 6 skills

Systems & MLOps

Databases, containers and compute.

InfluxDB
MongoDB
Docker
Linux shell
Git
KI:connect.nrw (AI clusters)

05 · 6 skills

Industrial domain

Where much of my data comes from.

Condition monitoring
Fault detection & diagnosis
Predictive maintenance
Industrial automation
PLC programming
Interoperability

§ 03

What I bring, with evidence.

Every strength below links to the project or publication that shows it. Filter by area, then open the evidence.

Methods
Condition assessment beyond fault detection
Estimating how much capability equipment retains, not only whether it has failed.
Methods
Physics-informed and data-driven modelling
Engineering knowledge and measured data are combined instead of treated separately.
Methods
Models that transfer between deployments
Reusable AI modules instead of one-off models for every site.
AI systems
Agentic AI for engineering workflows
Agents that coordinate real tasks such as model training and method deployment.
AI systems
Grounded natural-language interfaces
Answers that come from a system’s own data, not from a model’s general knowledge.
AI systems
Measured model quality
LLM systems are compared with defined metrics, not judged by impression.
Engineering practice
Grounding in industrial automation
Methods are designed by someone who has programmed the machines they describe.
Engineering practice
Communicating results
Results are presented to academic and industrial audiences.
Engineering practice
Careful data preparation
Sensor data is checked and cleaned before any model sees it.

§ 04

Selected projects.

From local language models and retrieval to time-series machine learning. Pick a project to read the problem, the approach and the outcome.

Project · Language · models · mapping1 / 9

Research · TH Köln (COLABO) · 2026

Local Language Models for Building-Data Mapping

Small language models, fine-tuned with LoRA and run locally, that map raw building-automation data points to digital-twin profiles inside an existing NLP pipeline, and do it more accurately than the hosted large model they replace.

83%F1 on unseen buildings

Problem

The pipeline that turns BACnet data points into Asset Administration Shells relied on a hosted 120B-parameter model: data leaves the site, runs are slow, and accuracy on unfamiliar buildings is limited.

Approach & contribution

Started with a reproducible experiment: LoRA fine-tuning of a small model to map point names to Brick classes lifted accuracy from 31 % to 97 % after 79 minutes on a laptop CPU, then exported to GGUF and served with Ollama
Fine-tuned 1.5B and 8B Qwen models (LoRA, GGUF quantisation) as drop-in replacements in the unchanged pipeline, trained on a GPU cluster and deployed on CPU
Curated a gold label set checked by a blind second annotator; curated labels turned out to be the key factor, since on LLM labels alone the small model only matched the LLM
Evaluated with grouped cross-validation across buildings, three seeds and McNemar tests

Outcome

On buildings not seen in training, mean F1 rose from about 67 % for the hosted 120B model to 77 % with the 1.5B model and 83 % with the 8B model; the 1.5B model runs on a laptop CPU in about half the time. Paper in preparation.

LoRAQwenllama.cppGGUFOllamaPyTorchSLURMBACnetAAS

§ 05

From production machines to AI systems.

2024 – nowResearch Associate (AI), TH KölnCologne, Germany · lived

Fig. B — Where I lived and where I travelled for talks. Select a stop.

§ 06

Publications

Peer-reviewed conference papers.

[1]
Ahmad, N., Krüttgen, M., Müller, J. (2026). Scalable Estimation of Functional Reserve in Technical Building Equipment Using Data-Driven and Physics-Informed Methods. AUTOMATION 2026, Baden-Baden. VDE Verlag, pp. 153–162.
First authorIEEE Xplore ↗
[2]
Krüttgen, M., Ahmad, N., Magsi, T. H., Müller, J. (2025). Automated Performance Evaluation of Buildings Based on a Taxonomy of Building Functions with NLP-Supported Exploration in Building Information Networks. CISBAT 2025, Lausanne. Journal of Physics: Conference Series 3140, 042018.
Co-authorDOI ↗
[3]
Krüttgen, M., Fessler, T. G., Magsi, T. H., Ahmad, N., Müller, J. (2025). Automatisierte Performancebewertung der Technischen Gebäudeausrüstung auf Basis einer ganzheitlichen Ordnungsmethode für Gebäudefunktionen. AUTOMATION 2025, Baden-Baden. pp. 471–482.
Co-authorIn GermanDOI ↗