Want to know if this job is worth applying to?
San Francisco, California, United States
Hybrid
About Pulley
Pulley helps the country’s top architects, builders, and retailers speed up every project in their portfolio. With AI-powered permitting intelligence and expert guidance, we eliminate costly delays and bring predictability across the full lifecycle of commercial projects.
Today, permitting is the slowest, most uncertain part of building, spread across 19,000+ jurisdictions with different rules, timelines, and surprises. Pulley gives project teams the clarity and predictability they need to move from planning to opening without delays.
We support rollout programs for brands like J.Crew, Solidcore, and Hibbett Sports, as well as major data center buildouts, EV charging networks, and other commercial projects. Our platform dramatically reduces approval timelines, improves forecasting accuracy, and removes thousands of hours of manual work from design and construction teams.
Founded in 2021, Pulley combines deep permitting expertise with purpose-built AI from people who have created products used by millions. We’re backed by CRV, Susa Ventures, Fifth Wall, and leaders from Plaid, Segment, ServiceTitan, and Procore.
In this role, you will build the intelligence behind the product that gets stuff built. Permitting runs on messy inputs—scanned plan sets, jurisdiction code, reviewer comments, application forms that differ in every city—and turning that into something fast, structured, and trustworthy is the core technical problem at Pulley. As a staff-level AI engineer, you will:
Own the AI problem space, not just features—define the technical direction for how Pulley applies LLMs across multiple product surfaces, and carry it from ambiguity through architecture to shipped, iterated-on product
Turn unstructured permitting documents, city regulations, and jurisdiction workflows into structured, reliable outputs—extraction, classification, retrieval, and agentic workflows over documents that were never designed to be machine-readable
Set the evaluation and observability standard for the company: decide how we define ground truth, measure quality and regressions, and know when a model change is actually an improvement—and build the systems that make that the default for every team shipping LLM features
Build with AI agents as a daily practice—directing, reviewing, and shipping agent-driven work at high velocity while owning the quality bar
Make the technical bets that determine what Pulley can build next year, not just this quarter—which models, which architectures, what we build versus buy—and own the consequences of those bets in production
Multiply the engineers around you: set the patterns others build LLM features within, mentor senior engineers toward larger scope, and make the whole team faster through the systems, standards, and abstractions you create
You thrive in ambiguity—you’d rather define the right problem than execute a spec, and you’re energized rather than paralyzed when the path isn’t laid out
You’re product-minded: you care whether the thing you built actually solved the customer’s problem, and you’ll talk to users to find out
You’re rigorous about what “working” means—you don’t trust a demo, you trust an eval, and you build the measurement before you build the feature
You have strong opinions about quality and velocity and don’t treat them as a tradeoff—you look for the tools, abstractions, and processes that buy both
You default to ownership at organizational scale: when something is broken or missing—a system, a process, a gap between teams—your instinct is to fix it, and you don’t need permission or a mandate to start
8+ years of software engineering experience, with a substantial portion building production LLM or ML systems
Track record of owning a significant AI domain end-to-end: from “this is somebody’s problem” through architecture, delivery, and production ownership, including the unglamorous parts—data quality, eval design, cost and latency, failure handling
Deep hands-on experience with large language models in production—prompting, retrieval-augmented generation, structured extraction, tool use and agentic workflows, and knowing when each is the wrong tool
Experience designing evals and otherwise making LLM-powered features reliable in production
Real experience building with AI coding agents—not just autocomplete; you’ve shipped work where agents did substantial implementation under your direction
Ability to architect durable systems while making pragmatic tradeoffs
Experience mentoring engineers or setting technical direction that other engineers built within
Based in the San Francisco Bay Area and willing to work in person 4 days a week
Experience with document understanding at scale—OCR, layout-aware parsing, or vision-language models over scanned PDFs, drawings, or forms
Experience fine-tuning models or building data pipelines to produce training and eval sets from real-world usage
Experience in construction tech, govtech, proptech, or another domain where the hard part is messy real-world documents and processes
Experience with modern full-stack development—we use TypeScript, React, and Google Cloud—and an appetite for working in the application code that puts AI features in front of users
Experience as the most senior AI engineer in a domain—being the person others escalated to when nobody knew the answer
Startup experience at the stage where you helped build the team, not just the product
