TensorRT Experts in Germany

• Developed and implemented AI-based perception and multimodal systems for real-world environments
• Built, trained, and evaluated Machine Learning and Deep Learning models using Python, PyTorch, TensorFlow, and OpenCV
• Worked with Vision-Language Models (VLMs), Large Language Models (LLMs), transformer-based architectures, and multimodal AI systems
• Applied LoRA-based fine-tuning techniques and experimented with diffusion models for generative and multimodal AI applications
• Developed multimodal perception pipelines using camera, LiDAR, and sensor data
• Designed end-to-end workflows for data processing, model training, evaluation, benchmarking, and robustness analysis
• Utilized HuggingFace Transformers and modern Deep Learning frameworks for AI experimentation and deployment workflows
• Applied GPU-accelerated computing, CUDA-based processing, ONNX, and TensorRT optimization for efficient inference and large-scale model training
• Collaborated with industry partners including Valeo and REHAU on applied AI and intelligent system projects
• Developed scalable AI architectures and prototype software solutions for automation and perception tasks

• Expert-supervised AI systems for scientific reasoning, model evaluation, and research workflows.
• Built the IAS Problem Solver, an orchestrated system for difficult mathematical reasoning; it achieved 84% in one submitted answer set on the Leipzig mathematics benchmark.
• Built a resumable state-machine pipeline for research-grade mathematics benchmark generation: source selection, LLM-agent-based phenomenon discovery, task synthesis, gold-answer and certificate generation and validation, probing, repair, human feedback, and quality gates, targeting tasks that are difficult, natural, verifiable, and cost-effective.
• Current work extends this into budget-aware AI research workflows for real scientific problems with expert review.

Tech stack: Python, OpenAI/OpenRouter-compatible APIs, embeddings, RAG, SQLite.

• Developed end-to-end reproducible ML pipelines (PyTorch) with data versioning (DVC), experiment tracking (MLflow), automated testing (PyTest), and CI/CD across all training workflows.
• Scaled Vision Transformer and CNN training across NVIDIA A100 GPU clusters (CUDA, DDP, SLURM); applied hyperparameter optimization (W&B Sweeps) to reduce training overhead and identify optimal configurations.
• Developed a real-time 3D human motion generation system (ViT, VQ-VAE, SMPL-X/PIXIE) for personality-conditioned avatar synthesis; achieved state-of-the-art FID = 6.15 and P-FID = 10.31 on the UDIVA benchmark.
• Validated model expressiveness through structured user studies, achieving 86% accuracy in distinguishing extroverted vs. introverted avatar behaviors.
• Optimized inference pipelines by deploying PyTorch models via TensorRT and ONNX Runtime into native C++ code; benchmarked performance.

• Architect end-to-end AI/ML solutions, with a strong focus on advanced LLM applications, autonomous agentic AI workflows using MCP toolset, vLM embedding computations, and robust agentic RAG pipelines to drive real-world business impact for clients across diverse sectors.
• Designed intelligent agentic AI workflows to automate knowledge retrieval and reasoning for B2B clients, accelerating business impact.

• Automated driver assessment by programming temporal rule engines to evaluate lane-change execution safety, head-pose mirror checks, indicator usage cycles, and compliance with traffic lights and road signs
• Synchronized real-time traffic sign recognition and multi-state traffic light classification models with time-series CAN-bus telemetry and HD-map spatial priors to grade traffic rule adherence
• Trained and deployed distinct deep learning models optimized for interior cabin monitoring and exterior surrounding-area perception
• Combined perception outputs with camera intrinsics and horizon stability checks to execute 3D ground-plane object distance estimation assuming flat-ground geometry
• Deployed a split-compute edge network across a 10-vehicle fleet via VPN, implementing a zero-allocation host memory pipeline to eliminate frame accumulation latency (6×21 FPS per vehicle)

Industry Software from Route to Invoice

In Production

In use at: landscaping company, two home care services

The entire workday, digital: sign-in, route plan on the vehicle, departure checks, photo documentation, time tracking, invoicing. The tablets keep working offline and sync afterward. Records are immutable, making time tracking easier to audit. Runs on a dedicated server in Germany with continuous monitoring and remote updates.

Proof: In daily use at the landscaping company with six vehicles since September 7, 2026.

• React Native on Android tablets
• Python
• SQLite
• Two-way synchronization
• Linux server in Germany
• Remote maintenance

• Developing methods for monocular 3D facial reconstruction and personalized geometric modelling from mobile imagery
• Building learning-based approaches for facial shape estimation, video-based facial analysis, and privacy-preserving visual learning

• Spearheaded the EU market entry strategy for L3/L4 autonomous logistics vehicles, driving the technological localization and deployment of the parent company’s smart robotics portfolio.
• Orchestrated technical alignment between top-tier autonomous driving suppliers across China and Europe, translating complex client requirements into precise engineering specifications compliant with EU standards.
• Cultivated strategic joint R&D initiatives with leading European universities, research institutes, and enterprises, accelerating the transition of cutting-edge robotic concepts into commercial products.
• Directed the end-to-end architecture of intelligent warehousing solutions, guiding cross-functional teams in optimizing hardware integration for autonomous vehicles & robots, and overall system performance.
• Led the R&D of high-fidelity simulation and AI algorithms using NVIDIA Isaac Sim & Lab, establishing robust "Sim-to-Real" pipelines to train and validate dynamic path planning optimization, intelligent obstacle avoidance, and complex navigation stacks prior to physical deployment.
• Maintained hands-on oversight of the core system architecture, focusing on bottom-level performance tuning, AI model inference acceleration with TensorRT/ONNX Runtime, and sensor integration.

• Held direct line management responsibility for a team of 4 Platform Engineers — owning hiring, performance reviews, and career development — while establishing a shared engineering standards framework and coaching culture that accelerated delivery across client engagements.
• Led a team of engineers to architect a cloud-native voice AI system for a major inspection client, enabling 2,500 field inspectors to document work fully hands-free via real-time transcription and AI agents — eliminating manual data entry across 440,000 inspections per month and reducing per-user cost from $9 to $1. Stack: AWS (DynamoDB, S3, Transcribe, CloudFront, API Gateway, Bedrock), ElevenLabs, Claude.
• Led a team of engineers to automate multi-region Kubernetes cluster management for a global SaaS leader, reducing provisioning time from 3 weeks to under a day and eliminating 90% of configuration errors. Stack: EKS, Terragrunt, Python, Bash, ArgoCD.
• Accelerator - Cloud-Agnostic AI Platform: Architected and delivered a cloud-agnostic, Kubernetes-native platform as an accelerator, enabling multi-tenant, enterprise-scale management of self-hosted LLMs with concurrent deployment of multiple base models and dynamic LoRA adapter serving. Designed production infrastructure using open-source tooling (ArgoCD, Karpenter, vLLM, SGLang) with automated model lifecycle management, API security (Keycloak + LiteLLM), and cost-optimized GPU provisioning.

• Topic: Object Detection and Semantic Segmentation for (AUV) Systems using Transformer-Based Vision Models and Sensor Fusion.
• Designed and implemented an end-to-end multi-sensor fusion perception pipeline (Camera, LiDAR, IMU) in ROS
• Developed CNN-based Machine Learning model (YOLOv8) and Transformer-based vision models for real-time object detection
• Processed and clustered 3D LiDAR point clouds using DBSCAN, RANSAC, and voxel grid filtering to enable robust object localisation in noisy environments.
• Designed Bayesian Network models (GeNle) for probabilistic reasoning and sensor-level decision fusion under uncertainty.
• Applied Kalman filtering for sensor state estimation, temporal alignment, and smooth object tracking, reducing false positives in safety-critical scenarios.
• Evaluated system performance under realistic driving dynamics, improving tracking stability and overall perception robustness.
• Built deep learning pipelines for training, validation, and performance evaluation of perception models using sensor data.

• Developed LLM-based automation utilities including structured reasoning pipelines, LLM-as-a-Judge evaluation tools, and multi-model comparison frameworks.
• Built RAG pipelines for internal research workflows using LangChain, ChromaDB, and FastAPI, enabling semantic retrieval and multi-step reasoning.
• Integrated LLM microservices into existing ML systems using Docker, FastAPI, and GitLab CI/CD with reproducible deployment workflows.
• Designed inference APIs combining vision models and LLM reasoning for multimodal analytics and decision-making.
• Optimized embedding-based retrieval using vector store pruning, improved chunking logic, and dynamic retriever selection.
• Performed prompt engineering and system instruction tuning for consistency, robustness, and reasoning quality.
• Built benchmarking suites to evaluate LLM latency, reasoning quality, retrieval accuracy, and robustness under different prompt templates.

• Deployed and optimized Vision-Language-Action (VLA) and diffusion policy models on NVIDIA Jetson Orin and Jetson Thor, meeting real-time inference latency targets for humanoid robot control loops.
• Built TensorRT engine pipelines (PyTorch → ONNX → TensorRT) with INT8/FP8 post-training quantization, calibration dataset design, and quantization-aware validation, reducing inference memory footprint by over 3× on Jetson without accuracy regression.
• Developed custom CUDA C++ plugins and CUDA Graphs for latency-deterministic, real-time policy execution – meeting hard runtime and memory constraints on embedded GPU targets.
• Developed an inference engine for VLA models on top of llama.cpp bringing different VLA policies under single runtime, packaging each as a single self-contained GGUF that needs no Python or PyTorch.
• Profiled and tuned GPU execution using NVIDIA Nsight Systems and Nsight Compute, identifying CUDA kernel bottlenecks, memory bandwidth saturation, and SM occupancy issues across Jetson Orin and Thor compute profiles for cross-layer performance optimization.