Hassan Rasheed: engineer, researcher, photographer

Hassan in a striped polo and sunglasses, smiling, in warm light

Hi, I’m Hassan. I studied computer science at NUST in Islamabad, shipped software in Lahore, then moved to Munich for a master’s at TUM and finished it with a thesis at Harvard Medical School. These days I live in Cologne.

My work sits where data engineering meets machine learning: pipelines that don’t break at 3 a.m., models that are useful outside a notebook, and lately agents that do real work for real teams. Off the keyboard I take photos, travel, and am slowly getting less bad at golf.

M.Sc. Data Engineering and Analytics: Technical University of Munich 2025
B.S. Computer Science: NUST, Islamabad 2019

Work

2026 – now
Forward Deployed Engineer at KI performance
Building agentic, multi-agent AI platforms for enterprise clients on Microsoft Azure.
Microsoft Foundry · Azure OpenAI · Azure AI Search · Python · MCP
Cologne
2024 – 2026
Software Engineer at peopleIX
Architected and owned the data and AI backend of an HR analytics platform: 50+ Airflow pipelines, multi-tenant Postgres models, a semantic metrics layer and the LLM stack behind its AI analyst.
Python · Airflow · AWS · Azure OpenAI · PostgreSQL · LangGraph
Cologne
2023 – 2024
Graduate Researcher at Harvard Medical School
Deep learning for tracker-less ultrasound guidance in neurosurgery: a texture-invariant keypoint descriptor that matches pre-op MR with intra-op ultrasound. Became my TUM thesis and a best-presentation paper.
PyTorch · Computer Vision · Contrastive learning
Boston
2023
Co-Founder at MuseLabs
A product mixing social media with e-commerce. Built the FastAPI backend on GCP and the Flutter app.
FastAPI · GCP · Flutter
Munich
2022 – 2023
Software Engineer, Cloud Processing at NavVis
Built the QA automation suite for release testing of the IVION Cloud Processing product.
Python · Selenium · Robot Framework · C#
Munich
2022
Research Assistant at Fraunhofer IKS
Constrained hierarchical reinforcement learning for safe robot navigation. Joint first-author paper at IEEE ICMLA.
PyTorch · Reinforcement learning
Munich
2020 – 2021
Machine Learning Engineer at SlashNext
ML pipelines for real-time phishing detection; a hash-based image method lifted detection on image data by 5%.
PySpark · TensorFlow · OpenCV
Lahore
2019 – 2020
Software Engineer at Techlogix
APIs and front-end for the web and mobile apps of a campus management product.
Angular · Ionic · Flask
Lahore

Research

Before agents and data platforms, I spent a few years in research labs. Most of it was one question: how do you match two images of the same thing when they look nothing alike? Along the way, a detour into keeping learning robots safe.

ASMUS @ MICCAI 2024First author

Best Presentation Award

Learning to Match 2D Keypoints Across Preoperative MR and Intraoperative Ultrasound

Hassan Rasheed, Reuben Dorent, Maximilian Fehrentz, Tina Kapur, William M. Wells III, Alexandra Golby, Sarah Frisken, Julia A. Schnabel, Nazim Haouchine

Method overview: (a) a stack of synthetic ultrasound images generated from one MRI slice, sharing the same keypoints; (b) a Siamese network with shared weights turning image patches into descriptor vectors trained with a contrastive loss; (c) keypoints on an MRI slice connected by match lines to the same points on an intraoperative ultrasound image. — (a) synthetic ultrasound generated from MRI, keypoints shared across the stack · (b) Siamese descriptor network trained with a contrastive loss · (c) matches between MRI and intraoperative ultrasound

The problem: During brain surgery the brain shifts, so the MRI taken beforehand stops lining up with reality. Ultrasound shows the current state, but it looks nothing like MRI: different texture, heavy speckle noise. Classic keypoint matchers find almost nothing in common between the two.
The idea: Matching by synthesis. From each patient’s MRI we generate many synthetic ultrasound images covering several MR contrasts and realistic ultrasound variation, so every keypoint has known correspondences. A Siamese network is then trained per patient with supervised contrastive learning to describe each keypoint by its structure, not its texture.
The result: On real clinical cases with ground truth from the ReMIND dataset, the descriptors reached 80.35% average matching precision, outperforming the state-of-the-art methods it was compared against.

My part Designed the descriptor and the synthesis pipeline, ran the experiments and wrote the paper. It became my TUM master’s thesis, done at Harvard Medical School / Brigham and Women’s Hospital.

Cite

@inproceedings{rasheed2024learning,
  title     = {Learning to Match 2D Keypoints Across Preoperative MR and Intraoperative Ultrasound},
  author    = {Rasheed, Hassan and Dorent, Reuben and Fehrentz, Maximilian and Kapur, Tina and Wells III, William M. and Golby, Alexandra and Frisken, Sarah and Schnabel, Julia A. and Haouchine, Nazim},
  booktitle = {Simplifying Medical Ultrasound (ASMUS 2024)},
  series    = {Lecture Notes in Computer Science},
  pages     = {78--87},
  publisher = {Springer},
  year      = {2024},
  doi       = {10.1007/978-3-031-73647-6_8}
}

MICCAI 2024Co-author

Intraoperative Registration by Cross-Modal Inverse Neural Rendering

Maximilian Fehrentz, Mohammad Farid Azampour, Reuben Dorent, Hassan Rasheed, Colin Galvin, Alexandra Golby, William M. Wells, Sarah Frisken, Nassir Navab, Nazim Haouchine

How do you find exactly where the surgical camera is, relative to the preoperative 3D scan, when the live image looks very different from the scan?

Split a neural radiance field into two parts: anatomy, learned before surgery, and appearance, controlled by a multi-style hypernetwork and adapted during surgery. The trained field becomes a differentiable renderer, and the camera pose is found by making its render match the intraoperative image.

“…the trained model is used as a differentiable rendering framework to estimate the surgical camera pose…”
Fehrentz et al., arXiv:2409.11983 (CC BY 4.0)

On retrospective clinical patient data it outperformed state-of-the-art registration methods while meeting current clinical standards.

My part Part of the same Harvard / BWH and TUM collaboration on cross-modal neurosurgical guidance.

Medical Image Computing and Computer Assisted Intervention, LNCS vol. 15006, pp. 317–327.

Cite

@inproceedings{fehrentz2024intraoperative,
  title     = {Intraoperative Registration by Cross-Modal Inverse Neural Rendering},
  author    = {Fehrentz, Maximilian and Azampour, Mohammad Farid and Dorent, Reuben and Rasheed, Hassan and Galvin, Colin and Golby, Alexandra and Wells, William M. and Frisken, Sarah and Navab, Nassir and Haouchine, Nazim},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI 2024)},
  series    = {Lecture Notes in Computer Science},
  volume    = {15006},
  pages     = {317--327},
  publisher = {Springer},
  year      = {2024},
  doi       = {10.1007/978-3-031-72089-5_30}
}

IEEE ICMLA 2022Joint first author

Safe Robot Navigation Using Constrained Hierarchical Reinforcement Learning

Felippe Schmoeller Roza, Hassan Rasheed, Karsten Roscher, Xiangyu Ning, Stephan Günnemann

Reinforcement learning agents learn by trial and error, which is a problem when an error means a robot crashing into something. How do you keep a navigating robot safe while it learns long, multi-step tasks?

Hierarchical RL splits navigation into a high-level planner that picks sub-goals and a low-level controller that reaches them. We added a safety layer that corrects the low-level actions so the robot stays inside a safe region.

In simulation, the constrained hierarchical agent kept its task performance better than a constrained flat agent while staying in the safe region.

My part Compared hierarchical RL methods and built the constrained hierarchical policy with the safety layer, at Fraunhofer IKS in Munich.

21st IEEE International Conference on Machine Learning and Applications, Nassau, Bahamas, pp. 737–742.

Cite

@inproceedings{roza2022safe,
  title     = {Safe Robot Navigation Using Constrained Hierarchical Reinforcement Learning},
  author    = {Schmoeller Roza, Felippe and Rasheed, Hassan and Roscher, Karsten and Ning, Xiangyu and G{\"u}nnemann, Stephan},
  booktitle = {21st IEEE International Conference on Machine Learning and Applications (ICMLA)},
  pages     = {737--742},
  publisher = {IEEE},
  year      = {2022},
  doi       = {10.1109/ICMLA55696.2022.00123}
}

Master’s thesis: Cross-Modal Medical Image Matching via Synthesis, TUM, carried out at Harvard Medical School / Brigham and Women’s Hospital (Surgical Planning Lab), 2024. Data: the public ReMIND brain-resection imaging database.