VLM Research Engineer (m/f/d)
Almetra · Onsite Berlin · Posted Dec 18, 2025
Transparency
DeadOnline for 48 weeks and never reposted. Jobs this old or reposted this often rarely lead to a hire. It may be filled already.
- First seen
- Oct 8, 2026
- Offered for
- 337 days
Overview
- Category
- AI & machine learning
- Employment type
- Permanent
- Workplace type
- On-site
- Country
- Germany
- Seniority
- Senior
- Schedule
- Full time
Key skills
- Vision-language model design
- Design and adapt multimodal models for video scene understanding and temporal reasoning.
- Large-scale GPU training pipelines
- Build and maintain distributed multi‑GPU training pipelines using PyTorch and mixed precision.
- Video‑text dataset curation
- Curate, augment, and generate synthetic labels for large video‑text and action datasets.
- Model compression and optimization
- Refactor architectures with pruning, quantisation, and distillation for efficient deployment.
- Production inference pipelines
- Deliver reliable inference services integrated with CV, platform, and robotics engineering teams.
- Benchmark development for video AI
- Create robust video QA, instruction‑following, and temporal understanding benchmarks to guide improvements.
Key phrases
- vision-language models
- real-world video understanding
- scene understanding
- temporal reasoning
- activity / action recognition
- large-scale training
- GPU clusters
- synthetic labels
- retrieval-based augmentation
- video QA
- pruning
- distillation
- production-ready inference pipelines
From the posting
We’re looking for a Research Engineer to push the limits of vision-language models for real-world video understanding. You’ll work on applied, state-of-the-art multimodal models and turn them into production pipelines used by customers.
Your role
Design and adapt vision-language and video models for scene understanding, temporal reasoning and activity / action recognition
Build and maintain large-scale training and evaluation pipelines on GPU clusters
Curate and augment video-text and action datasets, including synthetic labels and retrieval-based augmentation
Develop robust benchmarks for video QA, instruction following and temporal understanding, and use them to drive iterative model improvements
Cut and refactor model architectures for efficiency and deployability (compression, pruning, distillation)
Deliver production-ready inference pipelines to product and customer teams, working closely with CV, platform and robotics engineers
You bring
Completed PhD (or equivalent research track record) in computer vision, machine learning, robotics or a related field
Strong background in video-centric deep learning: scene understanding, temporal / activity / action recognition, or video generation
Experience training and adapting large vision or VLM models (e.g. InternVL, Qwen-VL, DeepSeek-VL, similar stacks)
Proven work with multi-GPU training (PyTorch, distributed, mixed precision) and large-scale datasets
Solid engineering habits: clean Python, reproducible experiments, reliable data and training pipelines
Track record of moving research into usable systems (demos, internal tools, or productised features) in fast-moving teams
Nice to have
Publications at top-tier venues (CVPR, ICCV, ECCV, NeurIPS, ICLR, etc.) on video, multimodal learning or scene understanding
Experience with 3D/4D scene representations, action generation or embodied / sense-plan-act style projects
Inference optimisation: quantisation, TensorRT, model distillation, or deployment on constrained hardware
Prior experience in a startup or applied research lab environment
What we offer
Employee Share Options Program for all permanent employees*
An increasing benefits list: currently includes Urban Sports club and quarterly team retreats.
Be on the forefront in defining what artificial intelligence means in manufacturing
Gain hands-on experience in working in an AI-first software company
Supportive and inclusive culture that values diversity and promotes the advancement of underrepresented groups within the company
Collaborate with a diverse (currently more than 10 nationalities) and talented team, working on cutting-edge projects with real-world impact
Network with professionals and leaders in the field, opening doors to potential future career opportunities
We have a very flat hierarchy, open 360° feedback, and flexible working hours
Ethics⚖: We are committed to developing ethical AI software
Don't meet all the requirements?
Almetra is committed to creating a workplace that is diverse, fair, and inclusive. We encourage candidates from all backgrounds, even if they do not meet every qualification, to submit their application. We firmly believe that having a team with diverse perspectives only strengthens our company and drives innovation. Our commitment also extends to providing an accessible environment for everyone, including those with disabilities. Please let us know if you require any accommodations during the application process or while working with us, and we will do our best to support you.
*Only full-time, permanent roles are eligible for stock options. Part-time roles, contract roles, work-student, internships and freelance roles are not eligible for stock options.
