Software Engineer, AI/ML

Interfere; New York (NY); On-site; USD 180,000 - 240,000; Full time

Benefits offered by this job

Health, dental, and vision

Visa sponsorship

Relocation to New York

Job summary

Interfere is building the operating system for product quality. You will own the intelligence layer that decides what is broken, why it happened, and what to do next, deploying scalable systems that run over real codebases and production data.

You will ship agents and LLM pipelines, enabling triage, root-cause analysis, and automated fixes at scale, with strong emphasis on measurability and continuous improvement.

Qualifications

• Shipped ML/LLM-powered systems in production with real users.
• Able to move between research and engineering and ship prototypes quickly.
• Capable of defining and delivering complete AI/ML solutions end-to-end.
• Treat evals as a core engineering problem with measurable impact.

Responsibilities

• Own the intelligence layer that powers Interfere's product quality platform.
• Build agents that reason over codebases, traces, and runtime data.
• Develop LLM pipelines for triage, root-cause analysis, and automatic fixes.
• Create detection systems for anomalies and data-driven feedback loops.

Skills

ML systems

LLM-powered

production experience

research-to-prod

eval-driven

agent architectures

multi-step reasoning

autonomous workflows

Tools

Agents

LLM pipelines

Evaluation systems

Retrieval & context

Job description

Interfere turns a product's invisible failures into shared problems the whole team can see and fix. Every app has places where users get confused, blocked, or forced to abandon a flow, but most of those moments never make it into a support ticket. We detect those failures in real-time and provide every person responsible for the fix with the context they need. When Interfere flags a broken checkout flow, the PM uses our data to prioritize the issue, the designer sees where the experience broke down, and the engineer pulls the trace underneath. We're building the operating system for product quality, so teams can move from scattered symptoms to a shared understanding of what's actually going wrong. We're a seven-person team in New York, with $5.1M raised from Y Combinator, Vercel Ventures, Hummingbird, Designer Fund, and others. Interfere is already running in production with design partners, which means the work you ship will immediately help real teams find and fix the failures costing them users today. The category is still being defined, but the product to fill this gap is inevitable, and the company that gets there first will own how the next decade of teams ship software. We're looking for the people who will move at the speed that demands.

The Role

About

You'll own the intelligence layer that makes Interfere work. The product is only as good as the systems that decide what's broken, what caused it, and what to do about it — and those systems live inside the agents, evals, and inference pipelines you'll build. You'll work where research meets production, shipping things that run at scale against real codebases and runtime data. Concretely, this looks like building:

• Agents that reason about codebases, traces, logs, and runtime behavior to find and explain bugs before users hit them, including the sandboxes they run in and the browsers and web research they use to investigate
• LLM pipelines for triage, root-cause analysis, and automated fix proposals, with the model routing that sends each job to the right model at the right cost
• Detection systems for anomalies in product behavior, performance, and user experience across noisy real-world data
• The evals, datasets, observability, and feedback loops that move us from "this prompt feels good" to "this system measurably works and is getting better," architecting the standards for the agentic parts of a product that is still being defined
• The context engineering that lets a model actually understand a customer's codebase: retrieval, indexing, context construction, and the integrations into their stack
• Systems that get sharper over time, turning every bug found and fixed into signal that improves the next detection, so continual learning is a product advantage rather than an afterthought

What we're looking for

• You've shipped ML or LLM-powered systems into production, where real users depend on the output, not just a notebook or a demo
• You move between research and engineering comfortably, and you pick up unfamiliar tech fast. You can go from a new framework or technique on Monday, to a working prototype by Friday, to an eval that tells you whether it's actually better the following week
• You can own something 0→1, beginning to end. You can take an ambiguous AI problem, define the next useful step, and ship without waiting for a fully specified plan
• You treat evals as a first‑class engineering problem. When the system is making decisions for users, intuition isn't a substitute for measurement
• Experience with agent architectures, tool use, multi‑step reasoning, or autonomous workflows in production

Nice to have

• Background in code understanding, program analysis, or systems that reason over source code
• Built retrieval, RAG, or context‑construction systems against large or messy data
• Anomaly detection, time‑series modeling, or learned monitoring systems experience
• Internalized understanding of AI/ML systems - a heightened understanding of the inner workings of the systems you'll be building

Strong signals

• An agentic system you built that other engineers actually chose to use, or that ran in production at meaningful scale
• Open‑source work, technical writing, or research with real depth — bonus if it shows where you disagree with the consensus
• You've shipped agents or LLM features into production and have informed opinions about what works, what's hype, and what's still broken
• You designed and ran evals that materially changed the trajectory of a product or research direction
• You started something from zero, whether a research direction, a product feature, or an internal tool
• You dropped into an unfamiliar domain, whether a new modality, a new codebase, or a new sub‑field, and shipped real work in it fast

How We Work

• We're in person in New York City. The hardest parts of building Interfere, from system design to architecture tradeoffs to taste calls on the product, happen faster and better at a whiteboard with people physically in the same room.
• We measure work, not hours. Time at a desk is a poor proxy for whether work is getting done. But there's a lot to do and genuine urgency to being the category winners. Most people who do well here end up putting in serious hours because the problems are interesting and the upside is real.
• We ship daily, and we ship deliberately. Speed and taste are not in tension here. Every line of code is a choice: we don't let tech debt accumulate because velocity is easier. We write the code we would want to inherit, while still pushing meaningful changes every day.
• The roadmap is structure, not scaffolding. We plan out the week, so there's structure to what you take on. But the items on that roadmap are whole features and subsystems, each one a project in itself. If you see another problem along the way that needs solving, you own that too. It's your job to make your work into what it needs to be.

Compensation and logistics

• Health, dental, and vision
• Visa sponsorship for exceptional international candidates
• We'll help you move to New York

Interview process

• Intro conversation with the founder
• Onsite/work trial with the team in New York
• References and offer