Full Job Description
Machine Learning Engineer, Platform
About the Role
So much of the work society depends on is still slower and harder than it should be. Permits take months. Claims sit unresolved. And AI hasn't changed that - because the bottleneck isn't the models. It's the institutional context AI needs to do the work: rules, history, relationships, and judgment scattered across people, documents, and legacy systems.
BrainCo exists to fix that. We build agent-native operating systems for the institutions society depends on, and our products are the first of their kind in the world - we were the first, anywhere, to fully automate construction permitting, and we're now doing the same across insurance and other industries. There is no playbook here, because no one has built this before.
As a Machine Learning Engineer on Platform, you'll build the core ML capabilities every product we ship stands on - built once, shared everywhere. This is the leverage seat in the company: improve document extraction, and every vertical improves with it; strengthen the blueprint foundation model, and every construction workflow gets sharper; ship a better improvement loop, and every system we've ever deployed keeps getting better on its own.
Come help build Atlas - our platform which includes a foundation model for the world's construction documents, extraction agents that read everything from policy stacks to financial filings, a unified eval system across every use case, model routing that puts the right model on the right task at the right cost. You'll own each capability end-to-end - from the pod that needs it this quarter to the abstraction that serves ten pods next year. Your customers are never abstract: they're the project pods building on your work, and through them, every institution we serve.
Who We're Looking For
You understand how machine learning actually works - not just the tooling, but the philosophy underneath: what a loss function really optimizes, how generalization breaks under distribution shift, why evaluation is where systems quietly go wrong. And you live at the bleeding edge of modern AI, with hard-won instincts for squeezing the most out of LLMs and agentic systems - prompting, fine-tuning, tool use, and reasoning. That combination is the job: you know when a fine-tuned segmentation model beats a VLM, when a rule engine beats both, and how to compose all three into a system more accurate than any single model. You treat frontier models as components to be measured, pushed, and engineered - never as magic.
You also have the platform instinct: you spot the general capability hiding inside three teams' specific requests - and know when generalizing is premature. You treat internal teams as real customers with real deadlines, and measure your success in their velocity.
Most of all, you're energized by building things that have never existed, and comfortable when the problem, the data, and the definition of success all have to be invented at once.
The Problems You'll Work On
Agents as shared capabilities. Document extraction, financial reporting, market data - built once, composed into many products. The dual bar: general enough for any pod to pick up, precise enough for decisions institutions stake their processes on.
Institutional Intelligence that compounds. Every verified correction improves the system twice: the corrected fact percolates to every application, and the system that builds the intelligence learns to build it better. You'll build the models behind both loops.
A foundation model for construction documents. The documents the built world runs on have never had a foundation model of their own. We have the data, the deployments, and the feedback loops to build one.
Model routing across every use case. The right model, at the right cost and latency, for every task - swapping frontier models underneath production systems without breaking institutional-grade guarantees.
One eval system for everything. A common language for quality across every use case - from segmentation models checking blueprints to agents adjudicating claims - that catches regressions before customers ever see them.
Composite AI systems and credit assignment. When a pipeline of vision models, VLM reasoning, and rule engines is wrong, which component failed? Because the components are shared, this is a platform problem - and one of the most interesting open problems in applied ML.
Continuous improvement, engineered. We promise customers their system gets measurably better every month it runs. You'll build the machinery that keeps that promise: capturing production corrections, triaging failures to the component that caused them, and turning that signal into retraining and safe redeployment - automatically, across every use case.
In This Role, You Will:
Turn pod needs into platform capabilities - find the general capability inside one team's specific, urgent request, without over-abstracting before the pattern is proven.
Own capabilities end-to-end. There is no handoff: whoever builds the capability owns its behavior in production, across every deployment that uses it.
Work at the research frontier with production stakes, turning LLMs, RL fine-tuning, and agentic systems into capabilities that dozens of institutional workflows depend on at once.
Serve customers on both sides of the wall - project pods as true customers, and when needed, the domain experts whose decisions your capabilities ultimately power.
Engineer for production reality, navigating accuracy, latency, cost, and reliability across environments far messier than any benchmark.
Raise the bar across the company. The platform is how learnings travel: what one pod discovers, you turn into something every pod inherits.
