Principal AI Engineer

The basics

• Your title will be Principal AI Engineer
• This role pays above market rate with significant equity.
• Matt, Cofounder and CTO, is the hiring manager
• We work Mon-Thu in our office in Chancery Lane, London, Fri from anywhere

What are we building?

Cursor built a $500m ARR business that predicts the next code change a software engineer wants to make. Why is there not a $500m ARR business that predicts the next email a salesperson wants to send? That's the business we've set out to build. The market is 10x bigger. Most of the modern economy is services - think real estate agents, recruiters, salespeople, insurance brokers, marketing agencies. And these people spend most of their time in email, doing admin: keeping their inbox organised, scheduling meetings, replying to repetitive emails, sending follow ups.

The goal is wide open. Big tech has failed to build anything useful for this type of customer, and there isn't a startup we know that's doing this that even has $1m ARR.

Since launching in April 2024, we've gone from $0 to $16m in ARR and raised a $30m Series B from top investors.

That's because we deeply understand the domain. Two of our cofounders, Rich and Archie, spent 8 years running an executive assistant agency, providing a manual service for what we're now automating using AI.

Email Is An Incredibly Hard Problem, And We've Spent The Past 18 Months Building All The Pieces We Need To Succeed

• Training data: We've assembled a world class human data division, composed of annotators that have experience at the world's top AI labs and model providers, to provide the training data for our models. And we've learned the playbook to get the most out of them
• Context: To predict the next email someone will send, you need to deeply understand their work. We've built battle tested, production grade integrations with our users' emails, meetings, documentation and messaging (slack etc) to do this. And have fine tuned models that read these and decide what makes it into our knowledge base.
• Feedback loop: No thumbs up thumbs down. We've developed an incredibly strong feedback loop that tells us whether a change we've made to our system works. Down to the word, and even character.
• Tools: There's lots of hype around generative AI. Different models coming out every day, new techniques. We've sorted the signal from the noise by experimenting with tools in production, to find what actually works. As a result, we know which base models to use, when and how to fine tune, when to use workflows vs agents, how to build the best RAG pipeline, etc. And we're refining our learning every day.

We're hiring a Staff AI Engineer to assemble all these foundations into something truly exceptional. Our goal is to predict the next email a salesperson will send 100% of the time.

What do we value?
Ownership
Each of our engineers owns a business area end to end. You'll be given a metric to optimise, and it's up to you to define the strategy to move that metric, as well as execute against it. We have no product managers and have no plans for that to change.

You'll be increasing the percentage of our draft content that gets sent.

Speed
We're very intentional about adding new people. We think a small team of exceptional people working hard at a problem they care about will always beat a larger, less focused team. That does mean you'll need to bring an intensity to this role that might not be asked at other companies. But it also means you will be fast tracked into more senior roles and responsibilities far earlier.

What will I do?
You'll Be Increasing The Percentage Of Our Draft Content That Gets Sent. Here's What We Think Will Move That Metric

• Jumping on calls with customers, and getting them to narrate their thoughts to you as they work through your email, giving you ideas for what we should build into the system to improve our performance
• Using the above to define what models we need to build, and how they should fit within the wider system
• Designing strategies to validate and troubleshoot the accuracy of each model step you design, as well as the accuracy of the whole system, and feedback the learnings to the product engineering team, and human data division. Whether using LLM evals or deterministic code.
• Working with product reliability engineers to improve the quality of our raw email, meeting, document and messaging data
• Architecting the structure of our knowledge base, how we save down to it, and how we retrieve from it
• Working with our human data team to get them to produce training and validation sets for those models, and checking data quality
• Working out where we need human data vs synthetic data

What does our ideal hire look like?

• You're truly obsessed with harnessing LLMs, agentic systems, and retrieval to achieve real world results, and have spent disproportionate time improving your understanding over the last 3 years, both in your day job, and in side projects.

• You have an expert level understanding of Typescript or Python, so that you can implement your ideas

• You have experience architecting and evaluating generative AI systems, working with some of the following:

• LLM evals

• Deterministic code based evals
• Vector database indexing and retrieval
• Fine tuning (supervised and direct preference optimisation)
• Interaction with human data teams

• Experimentation frameworks

• Urgency and intensity in your work. This is very hard problem, and will require disproportionate effort from you.

• Proactivity. This isn't a role where you'll be handed tickets. You'll be deciding how our AI system for predicting email replies develops alongside our CTO, with scope to change pretty much anything. You'll only be successful if you gain an understanding of the problem at a deep level by talking to customers and looking at individual data points, then continually update your knowledge on cutting edge approaches to applied generative AI.

Our tech stack
Broadly, we use a fairly typical serverless Typescript stack. It's not a requirement to have worked with every tool in this stack, but the more the better

• Typescript for all production code
• Pinecone as our vector database
• APIs into the models of all major providers (OpenAI, Gemini, Llama, Anthropic) for both fine tuning and inference
• Our custom human data and ML ops platform
• Vercel's agents SDK
• Firestore as our production database
• Firebase Auth as our auth system
• Backend deployed on Firebase Functions, and making use of PubSub and Cloud Storage
• BigQuery as our data warehouse
• Sentry and Google Cloud Logging for monitoring
• Github Actions for CI/CD

The application process

• Submit your CV (no need for a cover letter)
• An initial call with someone from the Fyxer AI team to review your experience and motivation for joining (15 mins)
• A call with our CTO to discuss your approach and experience (45 minutes)
• Live coding, remote (60 minutes)
• Meet team in office (60 minutes)