Mohammad Wasil Saleem's Portfolio

Mohammad Wasil Saleem

I'm a Machine Learning Engineer

Professional Summary

Machine Learning Research Engineer with 4 years of experience specializing in 3D Computer Vision, Neural Rendering, NLP, and Generative AI. My work involves dedicated research in Deep Learning, GenAI, Multimodal AI, and Reinforcement Learning, coupled with architecting production-scale MLOps workflows. My portfolio further explores the development of Games and 3D modeling.

Work Experience

Machine Learning Research Engineer - 3D Computer Vision & Neural Rendering

March 2026 - Present Deep Safety GmbH (DeepSafety.ai) Berlin, Germany

Neural Rendering: Trained 3D Gaussian Splatting (3DGS) and NeRF representations to perform photorealistic novel view synthesis from arbitrary camera positions, bypassing physical data collection bottlenecks and saving €10k+ per capture event.
To eliminate under-observed blind spots and volume artifacts, I rendered novel camera trajectories from base incomplete Gaussians, used generated video captions and camera conditioning to generate enhanced views via a conditioned video diffusion model, and distilled the inpainted frames back into a second Gaussian optimization stage.
Trained depth estimation architectures, introducing consistency-aware loss functions, alongside train and test-time augmentations to stress-test robustness and improve depth accuracy.
Orchestrated Structure-from-Motion (SfM) and point cloud registration workflows, utilized COLMAP for precise camera pose estimation and Open3D for depth fusion.
Converted 3D point clouds into high-fidelity meshes using Marching Cubes, Poisson Surface Reconstruction, and Nvblox, exporting metric-scale assets into Isaac Sim for simulation.
Verified point cloud alignment and 3D reconstruction fidelity using CloudCompare, Blender, COLMAP, and Isaac Sim.
Engineered robotic simulation environments in Isaac Sim by applying spatial transformations, dynamic rigid-body physics, and animated human agent trajectories, generating occupancy maps and automated multi-sensor (RGB-D) data collection pipelines for model training.
Extracted 3D human Gaussians from images using NVIDIA Asset Harvester for simulation.
Tech Stack: Python, PyTorch, Git, Bitbucket, Linux, Isaac Sim, Open3D, COLMAP, CloudCompare, Blender, SuperSplats, Nerfstudio

Machine Learning Software Engineer

April 2024 - Present Siemens Mobility Berlin, Germany

Siemens AI for Engineering : Project focused on the development and integration of AI solutions to automate domain-specific engineering workflows and significantly boost efficiency.
Artificial intelligence: transforming mobility for everyone: Company AI page
Designed and implemented Quality Agent, a company-wide tool for requirements analysis.
Developed a scalable multi-AI agent framework for automated test case generation for train braking systems.
Deployed across 7 distinct use cases, analyzing around 20 requirements per day.
Developed a self-service portal in JavaScript (GUI) and SQLite cache to improve adoption.
Integrated seamlessly with requirement management systems by developing secure APIs.
Engineered an automated Table Extraction pipeline (OCR) utilizing cloud-based AI solutions.
Designed to efficiently process highly unstructured, scanned technical documents.
Tech Stack: Python (LangGraph, Pydantic), JavaScript, SQL, Docker, Linux, AWS (IAM, S3, Bedrock, Textract), Azure (OpenAI, Document Intelligence), Git, CI/CD, Poetry, mypy, pre-commit.

Machine Learning Research Engineer - Computer Vision (Autonomous Trains)

March 2022 - February 2024 Siemens Mobility Berlin, Germany

Siemens Driverless Train : Project focused on Assistance and driverless train operations to maximize system capacity, improve safety, and ensure sustainability.
Assisted and driverless train operation: Company page
Developed and fine-tuned an end-to-end pedestrian detection model for autonomous trains using panoptic segmentation, aligning custom dataset with COCO dataset format.
Designed and implemented an Active Learning framework (uncertainty and diversity-based query strategies), boosting model accuracy by 89.94% across five cycles using just 7% of queried data, supporting ADAS safety.
Self-supervised Learning: Improved clustering of unlabeled image data by fine-tuning embeddings of ResNet-152 and VGG16 pre-trained with ImageNet dataset using contrastive learning and visualizing the results using t-SNE.
Built a semi-automated Data annotation pipeline using CVAT and models such as YOLO, SAM, MaskDINO, Mask2Former, and PanopticFCN, reducing the annotation time by 76%.
Leveraged Multimodal foundation models (BLIP and BLIP-2) for zero-shot transfer and semantic image retrieval to streamline the data curation process across large visual datasets.
Tech Stack: Python, PyTorch, Detectron2, CVAT, Git, Linux (WSL2).

Data Analyst - Research Assistant

February 2021 - March 2022 Leibniz-Institut für Agrartechnik und Bioökonomie e.V. (ATB) Potsdam, Germany

ATB Animal Welfare : Conducted statistical analysis of air exchange rates in naturally ventilated barns using methods such as ANOVA and model selection (AIC/BIC).
Individualized Livestock Production: Company page
Enhanced regression model performance by increasing R² from 0.65 to 0.85 through targeted feature transformation and optimization.
Developed scripts for task automation with Python and R; documented results in RMarkdown.
Tech Stack: R, Python, Excel, RMarkdown; Statistical and Probabilistic Modeling.

Unity3D Developer Intern

May 2018 - July 2018 ALG Media Private Limited New Delhi, India

Designed and developed the core mechanics and user interface for the Android 3D game.
Implemented dynamic runtime 3D mesh generation for tunnels, significantly boosting in-game performance.
Tech Stack: Unity3D, C#, JavaScript, Blender.

Education

Master of Science: Data Science

2019 - 2022 Universität Potsdam Potsdam, Germany

Grade: 1.8 GPA
Master's Thesis Topic: "Cost-Efficient and Model-Guided Multi-Query Strategy for Pedestrian Segmentation for ADAS in Railways"
Master's Thesis Grade: 1.1 GPA
In collaboration with Siemens Mobility
Thesis Link: Research Gate Link for Master Thesis
Proposal: Proposed an Active Learning-based Panoptic Segmentation method for railway safety, achieving an 89.94% AP improvement using only 7% of data points, while reducing labeling time by 76.19% through a deep learning–assisted human-in-the-loop annotation process.

Bachelor of Technology: Computer Science and Engineering

2015 - 2019 Jamia Hamdard New Delhi, India

Grade: 1.6 GPA
Bachelor's Final Project Topic: "Visual Question Answering"
Bachelor's Final Project Score: 463/500 (92.6%)

Skills

Languages

English (Fluent)

Deutsch

Hindi (Native)

Programming Languages

Python

JavaScript

SQL

Bash Script

Machine / Deep Learning Expertise / Computer Vision / LLMs

Neural Networks

CNNs

RNNs

Encoder-Decoder

Siamese Neural Networks

GANs

LLMs/SLMs

Transformers (ViT)

RAG / AI Agents

SFT / RLHF

PPO / DPO

PEFT / LoRA / QLoRA

Multimodal Models

Computer Vision

Neural Rendering

NLP

Generative AI

Image Generation & Inpainting

Novel View Synthesis

Active Learning

Contrastive Learning

Explainable AI (Grad-CAM)

Zero-Shot Transfer

Clustering & Dimensionality Reduction

3D Reconstruction & Neural Rendering

3D Gaussian Splatting (3DGS)

NeRFs

TSDF / SDF

Radiance Field Optimization

Point Cloud Registration

Voxel Grids

Mesh Generation (Marching Cubes)

Mesh Generation (Poisson Surface Reconstruction)

Mesh Generation (Nvblox)

Structure-from-Motion (SfM)

Camera Pose Estimation

SLAM

Python Libraries & Frameworks

PyTorch

Torchvision

NumPy

Pandas

Matplotlib

LangChain

LangGraph

OpenCV

Open3d

Nerfstudio

Hugging Face

NLTK

scikit-learn

MLflow

TensorBoard

WandB

pytest

Pydantic

Engines, Simulators & Asset Tools

NVIDIA Isaac Sim (MobilityGen)

NVIDIA Asset Harvester

COLMAP

CloudCompare

Blender

SuperSplats

MLOps / DevOps

Git

Linux

Docker

GitLab CI/CD

VectorDB (ChromaDB)

GitHub Actions

Cloud

AWS IAM

AWS S3

AWS SQS

AWS SNS

AWS Bedrock

AWS Textract

Azure OpenAI

Azure Document Intelligence

Blog Posts

Grad-CAM: What Does a Self-Driving CNN See Under Domain Shift? Blog

2024-2025

Portfolio

A comprehensive portfolio featuring Machine Learning and Deep Learning research, Reinforcement Learning, Neural Rendering and GenAI, supported by Docker-based engineering and Game Dev projects.

All
Machine/Deep Learning
Computer Vision / Neural Rendering
Natural Language Processing
GenAI
Reinforcement Learning
Game Development

Pedetrians Detection for Trains & Trams

Proposed an Active Learning-based Panoptic Segmentation method for railway safety, achieving an 89.94% AP improvement using only 7% of data points, while reducing labeling time by 76.19% through a deep learning–assisted human-in-the-loop annotation process.

Self-Driving Car (with Grad-CAM Analysis under Domain Shift)

Trained a Deep Learning-based autonomous driving model using CNNs for steering prediction and road line detection on Synthetic Data generated via Unity3D simulation, with real-time image streaming over TCP/IP. Used Grad-CAM to analyze how the model’s attention changes under different weather conditions, highlighting the effects of domain shift.

Neural Rendering for Autonomous Driving (WiP)

Built a Tiny NeRF (PyTorch) to render road scenes from car-mounted camera data, handling extrinsic calibration and coordinate alignment.

Neural Machine Translation with Attention

Implemented a Spanish-to-English Neural Machine Translation model with an Attention mechanism (from scratch), achieved a BLEU score of 25.37, comparable to the state-of-the-art.

Blender Text-to-Node ChatBot (WiP)

Developed an Open source 3D Blender add-on that translates natural language into Shader and Geometry nodes (text-to-code) by fine-tuning a Small Language Model (SLM) using RLHF with PPO and integrating it with Blender’s API.

Conditional GANs - Generate New Faces

Generated realistic faces from CelebA dataset using conditional GANs, optimizing attribute selection for improved results.

Predicting User Affiliation of YouTube Commenters with Hierarchical Attention

Built a large-scale dataset of 10 million YouTube comments and classified user affiliation using a Hierarchical Attention Network trained on 20K labeled samples, achieving 89.69% accuracy.

VQA: Visual Question Answering (Vision + Language)

Developed a multimodal (vision-language) model trained on the COCO dataset, achieving 51.28% accuracy on the Visual Question Answering (VQA) challenge.

Ticker Checker

Developed a full-stack, Docker-containerized intelligent support system that uses Deep Learning for ticket classification and integrates an optimized RAG pipeline with a Re-Ranking Algorithm to retrieve highly relevant historical tickets.

Q-learning vs SARSA applied to Smart Cab

Evaluated the efficacy of two Reinforcement Learning algorithms, Q-learning and SARSA, in the context of the Smart Cab game.

Take‘em Out (2018)

Developed an FPS Android game in Unity3D, featuring enemy AI, player health systems, and progressive level design for enhanced gameplay difficulty.

Contact

Reach out to me!

Email

wasilmohd1@gmail.com

Location

Potsdam, Germany