Skip to content

Computer Vision & Machine Learning Engineer
Würzburg, Bavaria, Germany

Open to full-time Machine Learning / Computer Vision roles in Germany

Faraz Kayani

I build Computer Vision systems that ship all the way to silicon.

Published at
CVPR Workshops
2026 · first author
Sensors
2026 · co-author
BMVC
2026 · accepted
Neurocomputing
2026 · under review
Portrait of Faraz Kayani
● edge detection · analysing…
LIVEfaraz@inference: ~
❯click to expand the console
read-only previewclick to expand ⤢
first author atCVPR Workshops 2026published inSensors 2026accepted atBMVC 2026under review atNeurocomputingresearch atCV Lab · CAIDAS & IFIbased inWürzburg, DEstatusopen to new rolescompletedM.Sc. Computer Science
01/ about

About

Computer vision and machine learning engineer. I design and train models, then make them small and fast enough for wherever they have to run, from a desktop GPU to a phone NPU. Architecture design, knowledge distillation, and quantization in PyTorch, measured on the hardware that will actually run them rather than estimated from FLOPs.

My first-author paper at the IEEE/CVF CVPR Workshops 2026 introduced LiteDenoiseNet, for which I wrote both the 41.6M-parameter teacher and the 1.96M student distilled from it. Around the models I built the lab's automated deployment and cross-device benchmarking pipeline, plus architecture deduplication for LLM-driven search. Two years shipping production software before the M.Sc. means the code holds up outside a notebook.

// email
faraz.kayani2322@gmail.com
// phone
+49 176 8503 8434
// linkedin
faraz-kayani
// location
Würzburg, DE
// languages
German A2 · English C1
// work permit
Full work permit · no sponsorship

Models

01

Training, distillation, and quantization in PyTorch. LiteDenoiseNet compressed a 41.6M-parameter teacher into 1.96M without giving up quality.

Systems

02

The engineering around the models: PyTorch to TFLite/ONNX conversion, ADB deployment to physical devices, cross-device benchmarking, and duplicate detection for LLM-generated architecture search.

Research

03

Four papers in 2026 came out of that work, first author on two. The code came first; the write-up followed.

02/ experience

Experience

A research lab now; before that, two years of agency deadlines. Both show up in how I work.

  1. run/0eaedaf

    Research Engineer, Computer Vision & Machine Learning

    Oct 2025 → Aug 2026

    Computer Vision Laboratory Würzburg · Würzburg, Germany

    • Designed and trained both the 41.6M-parameter teacher and the 1.96M-parameter student behind LiteDenoiseNet, reaching 37.58 dB PSNR at 34 ms per frame on a Dimensity 9500
    • Built an automated PyTorch to Android deployment pipeline that converts, deploys, and benchmarks a model on a physical handset with no manual step in between, open-sourced as nn-lite
    • Developed architecture deduplication for an LLM-driven neural architecture search system, stopping it from spending its budget regenerating models it had already produced
    • Traced a 3.88× NPU-over-GPU gap on identical models to operator support rather than compute, and turned it into the selection rule the lab now designs against
    • That work produced four papers in 2026 with faculty and PhD collaborators
    PythonPyTorchTFLiteONNXAI Edge TorchAndroid / ADBLLM toolingDistillationQuantizationC++Docker
  2. run/0007aae

    Software Engineer

    Jul 2021 → Jun 2023

    Zenkoders · Karachi, Pakistan

    • Delivered production web and mobile features end to end: REST APIs, database layers, and the React / React Native front ends on top
    • Integrated AI and chatbot functionality into live client products in Python and JavaScript
    • Shipped across several concurrent client engagements, scoping directly with designers, QA, and project managers
    PythonJavaScriptReactReact NativeREST APIsPostgreSQL
  3. run/17463b6

    Full-Stack Developer Intern

    Feb 2020 → Apr 2020

    Bitrupt · Karachi, Pakistan

    • Built React and React Native applications from prototype through to feature-complete delivery
    • Worked across JavaScript, Firebase, and SQLite in a small product team
    JavaScriptReactReact NativeFirebaseSQLite
03/ research

Research

full publication list on Google Scholar
★ first authorIEEE/CVF CVPR Workshops · 2026

Real Image Denoising with Knowledge Distillation for High-Performance Mobile NPUs

Faraz Kayani, Sarmad Kayani, Asad Ahmed, Radu Timofte, Dmitry Ignatov

Image DenoisingPyTorchKnowledge DistillationMobile GPU/NPUTFLite
Paper (PDF)↗NN-Dataset↗
co-authorSensors (MDPI) · 2026

Understanding the Performance of Deep Computer Vision Models: A Symbolic Regression Approach to Accuracy and Latency Prediction

Divyesh Rameshbhai Dhanani, Faraz Kayani, Saif U Din, Alice Arslanian, Dmitry Ignatov, Radu Timofte

Symbolic RegressionLatency PredictionModel AnalysisMobile Benchmarking
Paper↗
co-authorBritish Machine Vision Conference (BMVC) · 2026

Vision Model Inference on Mobile Devices: A Large-Scale Delegate-Aware Benchmark Beyond FLOPs

Saif U Din, Faraz Kayani, Radu Timofte, Dmitry Ignatov

BenchmarkingMobile InferenceDelegatesBeyond FLOPs
accepted · camera-ready submitted
★ first authorNeurocomputing · under review · 2026

Fourth manuscript on efficient vision models

Efficient Deep Learning
under review
04/ skills

Toolkit

Grouped by where in the stack it sits: from the training loop out to the interface.

// deep learning & vision

training, compressing, and analysing the models themselves

PythonPyTorchTensorFlowKerasOpenCVNumPyscikit-learnpandasHugging FaceJupyterKnowledge DistillationQuantizationNeural Architecture SearchLLM & RAG

// inference & deployment

getting a trained model onto real hardware, desktop or edge

ONNXTensorFlow LiteCUDA / GPUMobile NPUAndroid / ADBC++DockerLinuxBenchmarkingProfiling

// backend & data

the services a model has to live inside

FastAPINode.jsPostgreSQLFirebaseREST APIsCI / CDGit

// frontend & interfaces

where the predictions actually meet a person

ReactReact NativeNext.jsTypeScriptJavaScriptTailwind CSSExpo
05/ education

Education and certifications

M.Sc. Computer Science

Julius Maximilians University of Würzburg

2023 → 2026 · Würzburg, Germany

grade
1.8 · German scale
focus
Computer Vision & Machine Learning

B.Sc. Computer Science

DHA Suffa University

2017 → 2021 · Karachi, Pakistan

grade
1.3 · German scale
honours
Silver Medal

// credentials

  1. 01Machine Learning SpecializationDeepLearning.AI · Stanford Universitylink pending
  2. 02Generative AI with Large Language ModelsDeepLearning.AI · AWSverify ↗
  3. 03AI Skills FestMicrosoftverify ↗
  4. 04Network SecurityCiscoverify ↗
  5. 05CyberOps AssociateCiscoverify ↗

06 · contact

I train models.
And I ship them.

Looking for full-time machine learning and computer vision roles in Germany.

faraz.kayani2322@gmail.com→

or +49 176 8503 8434

LinkedIn↗GitHub↗Google Scholar↗CV↗