Senior Applied ML Researcher

About Ironsite

Construction is one of the most complex and labor-intensive industries, spending $7 trillion annually on labor, but productivity losses cost $1.6 trillion per year due to outdated management tools. Ironsite leverages wearable cameras combined with human labeling and AI vision language models to drive on-site productivity, safety & training for crafts workers. We put cameras on construction workers' hard hats and vests, then use advanced computer vision to analyze what's actually happening on job sites. We help teams reduce labor costs, improve safety, and deliver projects faster. To date, we’ve captured 50,000+ hours of construction footage across 7 states and have recently partnered with the nation’s #2 hard-hat manufacturer, Studson, to develop custom hardware purpose-built for the field. We’ve raised over $13M from 8VC, South Park Commons, and 30+ leading operators and technologists, including Eric Glyman (Ramp), Jeff Dean (Google), and Scott Wu (Cognition). Now, we’re building the team to scale nationwide. The Role

We're seeking an exceptional Staff / Principal Research Scientist to build our foundation model capabilities from the ground up. You'll work directly with our team to design, train, and deploy large-scale language and multimodal models tailored to construction domain expertise. This role offers the rare opportunity to train production LLMs from scratch using our unique, continuously-growing dataset of construction footage and operational data. What You'll Build

Foundation Models from Scratch : Design model architectures, training objectives, and optimization strategies to train large language models and multimodal models (combining vision + language) specifically for construction use cases Domain-Specific Pre-training : Develop pre-training strategies using construction documentation, safety protocols, building codes, and industry knowledge to create models that understand construction contexts deeply Multimodal Model Training : Build and train vision-language models that can reason about construction footage, technical drawings, and textual data simultaneously Training Infrastructure : Design and implement distributed training systems capable of handling billion+ parameter models across GPU clusters Model Evaluation & Alignment : Create comprehensive evaluation frameworks and fine-tuning pipelines (supervised fine-tuning, RLHF) to align models with construction domain requirements Data Pipeline Architecture : Build scalable data ingestion, cleaning, and preprocessing systems for training on diverse construction data sources Production ML Systems : Deploy and monitor computer vision models in challenging real-world environments with strict latency and reliability requirements Human-AI Collaboration Platform : Develop systems that seamlessly integrate human labeling with AI predictions for continuous model improvement Cross-modal Understanding : Create models that combine visual data with contextual information for deeper construction insights

Technical Challenges You'll Solve

Training large-scale models efficiently with limited compute budgets while maximizing performance Developing novel pre-training objectives that capture construction-specific knowledge and temporal reasoning Implementing efficient attention mechanisms and architectural innovations for long-context understanding of construction projects Designing evaluation metrics that measure real-world construction task performance beyond standard benchmarks Balancing model capability with deployment constraints for edge and mobile applications

What We're Looking For

Technical Excellence [REQUIRED] 5+ years of experience in deep learning research with focus on large-scale model training [REQUIRED] Demonstrated experience training models at scale (100M+ parameters) from scratch [REQUIRED] Deep understanding of transformer architectures, attention mechanisms, and modern training techniques (mixed precision, distributed training, gradient accumulation) Production experience with ML systems at scale (PyTorch, distributed training, model serving) Experience with language model pre-training, including tokenization strategies, training objectives (CLM, MLM), and scaling laws Understanding of fine-tuning techniques including instruction tuning, RLHF, and preference optimization (DPO, PPO) Knowledge of efficient training techniques: LoRA, QLoRA, flash attention, gradient checkpointing Experience with model evaluation, benchmarking, and safety considerations Systems Experience Background in vision-language models or multimodal architectures strongly preferred Experience building and optimizing data pipelines for large-scale training Systems-level thinking about training efficiency, hardware utilization, and cost optimization Familiarity with MLOps for model versioning, experiment tracking, and reproducibility

Preferred Qualifications

Advanced degree (M.S./Ph.D.) in Computer Science, Electrical Engineering, or related field with focus on deep learning First-author publications at top ML venues (NeurIPS, ICML, ICLR, ACL, CVPR) on language models, multimodal learning, or efficient training Experience training or contributing to open-source foundation models Background in domain-specific model development (code, science, medical, etc.) Experience with video understanding models or temporal reasoning in transformers Contributions to major ML frameworks or training libraries Track record of transitioning research prototypes to production systems

Location & Compensation San Francisco Bay Area (on-site) Competitive salary and significant equity package Full benefits including health, dental, vision, and 401k +6% match Access to dedicated GPU compute resources for research and experimentation

Compensation

The base pay range for this role is $180,000 – $350,000 per year.