About ZeroDrift
ZeroDrift is building the AI-native compliance enforcement infrastructure for enterprise communication. We are the first platform that enforces the rules before an AI message is ever sent, fixing violations in real time across every channel where AI speaks for the business. Everything goes out clean. Nothing dangerous comes in.
As AI starts to communicate on behalf of entire organizations, the gap between what compliance requires and what companies can actually enforce is widening fast. Every existing solution monitors after send, once the risk is already out the door. ZeroDrift enforces before send. Enforcement did not exist as a category. We created it.
Backed by Andreessen Horowitz (a16z speedrun), PitchDrive Ventures, Reign Ventures, and other leading investors, ZeroDrift is built by engineering and AI leaders from Microsoft, Google, and Goldman Sachs, led by a repeat AI founder.
The role
You own the model at the core of ZeroDrift: the small, purpose-built language model that decides, in real time, whether a message is safe to send, and rewrites it when it is not. As our Principal Applied Scientist, you are the most senior individual contributor on our AI, and your mission is to make this the most accurate compliance model in the world, at low cost and low latency. This is a hands-on role. You build and own the model yourself, you do not manage a team. You will be the first and most senior person on the model, working directly with the founder, with the ownership and equity upside that come with defining a category from day one.
What you will own
- Own the compliance detection and rewrite model end to end: data, architecture, training, evaluation, and what runs in production
- Build the evaluation harness and the labeled-data flywheel that prove how good the model is. This is your first and highest-leverage priority, because we cannot claim best in the world without it
- Own accuracy, cost, and latency, and the tradeoffs between them
- Own the model architecture direction. Today the detection and rewrite logic runs on a frontier model alongside a retrieval and rules layer. You own the path to a smaller, specialized model that matches or beats it on accuracy at a fraction of the cost and latency, and you decide how and when we get there
- Build the infrastructure that adapts the model to each customer's rules and policies, including adapters and per-tenant tuning
- Turn plain-language policy into model behavior the system can enforce, and own the reasoning behind every decision the model makes
- Run tight error analysis on the cases that matter most: violations that should have been caught, and false alarms that erode trust
What you bring
- 8+ years in applied ML or AI, with models you have shipped and kept healthy in production
- Hands-on depth in modern NLP and LLMs, including fine-tuning, distillation, and rigorous evaluation
- The judgment of a senior IC: you set technical direction and define what good means, not just execute it
- Strong instincts for model design tradeoffs: when to fine-tune versus prompt, when to distill, and when a smaller model is good enough to ship
- A bias for shipping, and for measuring everything you ship
Bonus
- Experience taking a system from large general models to smaller, specialized ones for cost and latency
- Compliance, regtech, or another regulated domain
- Experience adapting one model to many customers (adapters, per-tenant tuning)
- Public research, open source, or technical writing
How we work
We are hyper-focused on speed of execution and on owning our work end to end. This is a full-time hybrid role: three days a week in our office at 4 World Trade Center in New York City, and the rest remote.
Benefits
Top-tier medical, dental, and vision coverage, plus a 401(k).
ZeroDrift is an equal opportunity employer. You must be authorized to work in the United States.
