Ahmed Bahnasy

AI Engineer / MLOps Engineer - APTIV (Automotive)

Tech Lead — In-Vehicle AI Agents (Nov 2025 - present)

• Led the design and development of the Intelligent Cockpit Assistant (VOS) — an Emotional Cockpit that senses its occupants, adapts to their needs, and enhances safety, comfort, and engagement through a voice-first AI assistant with a growing catalogue of specialised in-car services accessible through natural conversation
• Built a full-stack AI solution combining multimodal sensing — in-cabin cameras, vehicle state, GPS, and surrounding cameras — with generative AI and contextual content delivery, so the cockpit understands who is inside, how they feel, and what is happening around the car
• Built the multi-agent orchestration layer over MCP / A2A: 13+ specialised agents (vehicle control, DMS, RAG, web search, vision, translator, app store) coordinated by a launcher agent supporting both delegation ("ask") and full handoff ("launch"); fluid back-and-forth dialogue with barge-in / VAD pipelines for sub-second turn-taking
• Made the assistant pluggable across cloud and on-device LLMs/VLMs (Bedrock, Gemini, vLLM, OpenAI-compatible) with a hybrid edge–cloud inference topology, and took the system from prototype to working product demo running on real car hardware

Tech Lead — MLOps / ML Infra (Apr 2023 - Nov 2025)

• Built the team's cloud ML infrastructure from scratch — hybrid on-prem + AWS EKS Kubernetes with Terraform and Helm, GitOps via ArgoCD, Karpenter for GPU autoscaling — and managed the 204K/year compute and storage budget
• Built Airflow pipelines for automated dataset preparation, model training, and automated evaluation — metrics, ablation tracking, and experiment comparison integrated with MLflow
• Full observability stack for live training and serving: Prometheus, Grafana, ELK with GPU and system-level metrics
• Managed terabyte-scale CV dataset lifecycle (acquisition, annotation, versioning with DVC, validation) ensuring data quality, traceability, and reproducibility
• Optimised distributed training on SLURM and Kubeflow clusters: multi-GPU DDP, GPU scheduling and resource allocation, and profiling for faster, reproducible training cycles

Master Thesis Student - Helmholtz Zentrum München

• Thesis: "Optimization of instance segmentation performance for fragmented and strong occlusions in neonatal incubator environments"
• Research collaboration resulted in publication on wireless colorimetric multi-biomarker sensing — accepted at HAICON 2026 (preprint on medRxiv, 2025)

Working Student - ESR Labs, Accenture

• Developed deep learning algorithms for camera-centric automotive perception
• Worked on Monocular Depth Estimation and Segmentation tasks
• Training, evaluating, and optimizing inference time of DL models

Student Researcher - Visual Computing Lab, TUM

• Guided Research Project on 3D object detection and tracking using Graph Neural Networks for Autonomous Driving

Working Student - Allianz SE

• Developed statistical learning models for customer analytics and communicated insights to stakeholders

Software Engineer - Valeo (Automotive)

• Developed AUTOSAR/MISRA-C compliant SW components for ECU diagnostics
• Built CANoe/CAPL test environment reducing release time by 80%
• Projects: Peugeot S.A. Front Camera, Climate Control, headlights, and steering ECUs

Control Engineer - Advansys, ESC

• Design and run test scenarios for PLC control software

M.Sc. Informatics - Technische Universität München (TUM)

• Focus: Machine Learning and Computer Vision
• Coursework: Multiple View Geometry, Tracking and Detection in CV, Advanced Deep Learning in CV, Machine Learning in CV

B.Sc. Mechatronics Engineering - German University in Cairo

• Semester Abroad: Berlin Campus (Sep 2014 - Jan 2015)
• Coursework: Image Processing, Visual Servoing, Autonomous Systems, Data Structures & Algorithms

VOS — Voice Operating System for Vehicles (2025-2026)

• Voice-first in-vehicle AI agent platform with on-device VLM, multi-agent orchestration over MCP / A2A, and edge deployment topology (board + UPU)
• 13+ specialised agents (vehicle control, DMS, RAG, web search, vision, translator, app store) coordinated by a launcher agent supporting both delegation ("ask") and full handoff ("launch")
• Multi-provider LLM executor (Bedrock, Gemini, vLLM, OpenAI-compatible); barge-in / VAD pipelines and streaming SSE for sub-second turn-taking
• Read more →

On-premise VLMs/LLMs Inference Platform (2026)

• Designed cost-optimized, scale-to-zero LLM/VLM inference platform on EKS
• Implemented OpenAI-compatible API and ChatGPT-like web interface (Open WebUI)
• Configured Karpenter for GPU node provisioning with Spot instances, reducing idle costs by 70%
• Built model routing service with request status tracking, Redis queuing
• Enabled multi-model serving (7B-70B+ parameters) with EFS-backed shared model storage

Cloud ML Infrastructure from Scratch (2024-2025)

• Designed and provisioned hybrid cloud infrastructure (AWS + Azure) using Terraform
• Built EKS cluster with Karpenter for GPU autoscaling and Kubeflow for distributed training orchestration
• Implemented GitOps workflows with ArgoCD, deployed ELK stack for logging
• Set up Prometheus and Grafana for system and GPU metrics monitoring

Kubernetes Extension for Hardware-in-the-Loop Testing (2025)

• Developed custom CRD and Device Plugin integrating embedded target boards as schedulable Kubernetes resources
• Enabled HIL tests for embedded software within the cluster

BlenderProc (2020)

• Contributed the first human pose data loader using the AMASS motion dataset for photorealistic training image generation