MTS - Post-Training Lead

About Ironsite

The Role

Open Problems You Could Own in Your First Year

• Post-training for expert-level perception. Using SFT and RL (GRPO and beyond) to push VLM labeling of fine-grained construction activity beyond expert human taggers, across every trade
• RLVR for agents that investigate the jobsite. Training agents with verifiable rewards to go beyond labeling and run their own research loop over a site's footage: form a hypothesis about what's slowing a crew down, dispatch a swarm to search thousands of hours of video for evidence, and return recommendations field leaders can act on the next morning
• Long-video understanding at industrial scale. Reasoning over multi-hour egocentric footage where the events that matter are sparse. This requires temporal grounding, long-context modeling, and memory beyond what current VLMs offer
• From understanding to prediction. Moving from models that describe what happened on a jobsite to models that forecast what happens next: how a crew's current sequence plays out over the coming days, where work will stall, and what changes would prevent it. The first step toward a true world model of construction
• Inference at fleet scale. In the next twelve months we'll collect and process millions of hours of video per month, eventually hundreds of thousands per day. Making frontier-quality inference cheap enough to run on all of it (distillation, quantization, model routing, hardware-aware optimization) is a partially unsolved problem you'd own
• Evaluation frameworks that predict reality. Building benchmarks that track real field performance as we expand to new jobsites, new trades, and even the same jobsite at a different stage of work. Standard academic benchmarks don't cut it for what we're solving

What You'll Do

• Architect and train novel VLMs. Design, train, and iterate on general-purpose vision-language models fine-tuned for spatial intelligence in construction environments. Set the bar for model quality across the research team
• Own the post-training roadmap. Take the lead on executing our research goals, from establishing baselines with state-of-the-art models to developing post-training recipes (SFT and RL), long-context architectures, and visual reasoning techniques
• Build on the Construction Intelligence Benchmark. Expand our existing benchmark suite across video question answering, temporal reasoning, activity recognition, and site-level analytical reasoning. Make it the reference point future construction AI research measures against
• Build scalable pipelines. Develop and own the model training and evaluation pipelines. Ensure we can rapidly experiment, measure performance, and deploy models into production
• Ship at production scale. Apply distillation, quantization, and model routing so state-of-the-art understanding runs affordably across millions of hours of monthly footage. Partner deeply with hardware and infrastructure teams so research decisions and deployment realities inform each other from day one
• Provide technical leadership. Set the technical direction for post-training, review designs and PRs, mentor researchers and interns, and set the standard for experimental rigor. This is technical leadership, not people management (though you'll be central to hiring the next generation of researchers)

Technical Challenges You'll Solve

• Training frontier models efficiently under real compute budgets while maximizing performance on the problems that matter for our customers
• Inventing new post-training objectives that capture construction-specific knowledge, temporal reasoning, and fine-grained perception at a level generic approaches can't reach
• Designing efficient attention mechanisms and architectural innovations for long-context understanding of construction workflows that span hours, days, and multi-project trajectories
• Building evaluation frameworks that measure real-world construction task performance beyond standard benchmarks, and defining what "good" means for a category the field hasn't yet formalized
• Balancing model capability with deployment constraints for edge, on-prem, and cloud inference across a fleet growing by orders of magnitude
• Working at the seams between research and production. The most interesting problems at Ironsite live where a training decision propagates all the way through to hardware constraints on a jobsite in Texas

What We're Looking For

Required

• Deep experience owning post-training (SFT and RL) for a large language or vision-language model that shipped to production. You've done this at least once, ideally more than once
• A track record of research contributions that meaningfully advanced the state of the art through papers, models, systems, or products others in the field have built on
• Deep expertise with modern deep learning frameworks (PyTorch, JAX, or similar) and strong proficiency in Python with solid software engineering fundamentals
• Deep experience working with and creating large-scale vision or language datasets
• Comfort operating at the frontier of what's known. You've led research on problems where the right answer wasn't in a paper yet, and you figured it out anyway
• Experience mentoring senior researchers and setting technical direction for a team
• A background in Computer Science, Machine Learning, AI, Robotics, or a related technical field, or the equivalent hands-on experience

Strongly preferred

• Deep expertise in fine-tuning and post-training large language or vision-language models (SFT, GRPO and other RL methods, parameter-efficient tuning such as LoRA, and novel post-training recipes)
• Hands-on experience with the hardest challenges of video data, including temporal reasoning, long-context modeling, memory, and efficient processing at scale
• Experience optimizing inference at industrial scale, including quantization, distillation, sparsity, and efficient serving on edge and cloud infrastructure
• Strong publication record at top-tier AI, ML, or CV conferences, or comparable evidence of research impact (open source, product influence, community leadership)
• Prior experience at a frontier lab, applied AI startup at scale, or research-heavy product company where you shipped post-training work to production

Nice to have

• Familiarity with MLOps tools for scalable model training and deployment
• Experience with multimodal models spanning vision, language, and additional modalities (audio, sensor, motion)
• Strong interest in vision-language models applied to real-world physical problems, and genuine curiosity about the day-to-day lives of construction workers

What Success Looks Like

• First 30 days: You know our data, our benchmarks, and our production models cold. You've identified the highest-leverage post-training directions for the next twelve months and made your case for what Ironsite should invest in and why
• First 3 months: You've led at least one major post-training initiative from problem definition through deployment. Your work has visibly moved the state of Ironsite's model performance on a problem that matters, and you're actively shaping the research roadmap alongside the Chief Science Officer
• First 6 months: You're one of the defining research voices at the company. You've mentored the rest of the research team, hired at least one senior researcher, and led the technical direction on a piece of Ironsite's post-training strategy that will run for years

Location, Compensation, & Perks

• San Francisco Bay Area (on-site)
• Base salary: $250k-$400k per year, commensurate with experience
• Significant early-stage equity
• Full benefits including health, dental, vision, and 401(k) with 6% match
• Access to dedicated GPU compute resources for research and experimentation
• Daily catered breakfast and lunch
• Office in San Francisco, next to Oracle Park and the Caltrain

Why You'll Love Working at Ironsite

• Foundational impact. Solve fundamental AI problems to transform one of the world's largest and least-digitized industries. Your models ship to real jobsites, not just papers
• Ownership and autonomy. As Post Training Lead, you'll have real authority over the post-training direction, technical decisions, and how the team operates. We value intellectual curiosity, first-principles thinking, and iterating quickly to turn ambitious ideas into reality
• Dream dataset. Exclusive access to a massive, proprietary, and continuously growing corpus of egocentric jobsite video from hundreds of devices deployed on active construction sites. A moat that enables frontier research no other lab can do
• World-class team. Lead alongside a small, elite team of researchers and engineers who have shipped cutting-edge AI products at companies like DeepMind, Etched, Meta, Apple, and NVIDIA