Own the system architecture for hybrid (cloud/edge) voice assistant systems — define the edge/cloud allocation model, routing criteria, and the port/adapter boundaries that allow functions to move between cloud and vehicle without redesign.

Design and oversee end-to-end prototype pipelines (ASR → LLM → TTS), including integration into early vehicle platforms (E³, SDV) and test operation in lab and driving environments.

Drive evaluation and benchmarking of base components: on-device ASR for edge hardware, modern TTS models (latency, robustness, audio quality, energy efficiency), and candidate LLMs for dialogue-based assistant functions via Azure AI Foundry.

Define RAG architectures for improved knowledge coverage and robustness, including embedding strategies and small-scale vector stores for specific automotive use cases.

Architect the embedded/edge AI optimisation workstream: quantisation, pruning and distillation for ARM/DSP/NPU targets, with explicit attention to memory footprint, energy consumption, wake-word efficiency and thermal behaviour.

Shape the research agenda for future key technologies — continual learning and on-device model adaptation (anti-drift), multimodal interaction (speech + image + vehicle sensor data), emotion and sentiment recognition, privacy-compliant on-device personalisation and long-term memory, multi-speaker handling, speaker identification and anti-spoofing.

Embed safety, security and privacy mechanisms into concepts from the start (differential privacy, secure enclaves, automotive security standards) and ensure interoperability with vehicle architectures (SOA, microservices, zonal architecture).

Apply and enforce hexagonal architecture as the structural standard, so prototypes are transferable into series development rather than thrown away.

Produce the technical concepts, architecture decision records and evaluation reports that hand research results over to the series development work package — and defend them in review with the client's architects.

Lead the offshore engineering team as Location Lead: technical direction for AI/ML and AI Ops engineers, code and design review standards, work breakdown and estimation across a high-throughput ticket flow (~30 tickets/sprint, sizes S/M/L), onboarding and skills growth.

Act as the offshore technical counterpart to the onshore solution architect and client engineering teams; run technical alignment across time zones and represent the location in architecture boards and sprint ceremonies.

Edge / on-device AI: model optimisation (quantisation, pruning, distillation) and deployment to constrained hardware — ARM, DSP, NPU, or mobile/embedded equivalents.

Expert Python.

Hands-on with LLM frameworks in production: LangChain, LangGraph or equivalent.

3+ years as architect or technical lead, with end-to-end ownership of a system's design (within 8+ years total engineering experience).

Track record of taking AI/ML or LLM systems from prototype into production — not research-only.

Practical experience building RAG systems: embedding models, vector stores, retrieval evaluation.

Azure OpenAI Service or OpenAI API in practice, plus Azure cloud (AKS or Container Apps).

Integration experience with at least one ASR or TTS engine (Whisper, Azure Speech, Cerence or equivalent).

Experience leading a distributed engineering team (5–10 engineers) in Scrum/Kanban.

English C1, and willingness to travel to Germany for workshops, architecture reviews and prototype test campaigns.

Hexagonal / ports-and-adapters architecture and API design (REST, gRPC, protobuf, OAuth2) — the client mandates this architecture style, so the candidate must be willing to adopt it quickly.

Automotive / in-vehicle software context: infotainment platforms (MIB3, E³, SDV), Android Automotive / AAOS, AIDL, Viwi, vehicle telemetry services, SOA or zonal E/E architectures.

In-vehicle speech processing specifics: noise cancellation, far-field microphones, wake-word engines, barge-in, multi-speaker cabin scenarios.

Edge inference toolchains: ONNX Runtime, TensorRT, TFLite, OpenVINO; audio signal processing with pyAudio / librosa / pydub.

Observability and evaluation for LLM systems: LangFuse, OpenTelemetry, and evaluation frameworks (DeepEval, Ragas, PromptFlow); quality metrics such as BLEU, WER, CER, faithfulness and hallucination rate.

Commercial automotive speech/emotion stacks (Cerence, HumeAI) and Google TTS.

Privacy-preserving ML: differential privacy, federated/aggregate learning signals, secure enclaves, on-device personalisation.

Speaker identification and anti-spoofing.

Data/persistence breadth: PostgreSQL + pgvector, MongoDB, CosmosDB; IaC with Terraform/Terragrunt.

Awareness of automotive quality and security frameworks (A-SPICE, ISO/SAE 21434, TISAX) and of EU AI Act implications for AI systems.

German language skills.