Senior Backend Engineer

We're hiring a senior engineer for the hardest system we run: real-time infrastructure in the path of the live traffic of AI coding agents, deployed inside our customers' own clouds, where a mistake is visible to a developer within seconds. Node.js is the tool; the job is protocols, telemetry, endpoints, enterprise operations and the economics of AI, with the product side of your own work on top.

About ZuzAI

AI coding agents and assistants are changing how software is built, and they are becoming one of the fastest-growing, least-understood lines on an enterprise's bill. Engineering leaders can't yet answer the basic questions: what it costs, what it produces, and whether it's being used the way the company intended. ZuzAI builds the infrastructure that answers them. We're deployed in production at global technology companies, inside their own clouds, and we ship to them every week.

We're a small, senior, agent-native engineering team, on-site (3 days a week) , funded, and early enough that what you build defines the product. You'd work directly with the Director of Engineering and have a direct line to the founders.

01 The challenge

AI coding agents and assistants are the fastest-changing software ecosystem there is, and large organisations are adopting them faster than they can understand what they cost, what they produce, and who is allowed to do what. ZuzAI builds the infrastructure that gives them control. The data plane is the part that lives on the traffic itself, and it is hard in ways a typical backend isn't:

• Protocols nobody documented. Each client speaks its own dialect over HTTP/2, WebSocket, protobuf and server-sent events, and changes it without notice, sometimes weekly. We read it as it passes, with a packet capture on one screen and production traffic on the other.
• From traffic to truth. A pipeline that turns raw streams into one taxonomy of sessions, turns, tool calls, shell commands, skills and code evidence, with identity attached and every token priced the way the vendor bills it.
• AI provider knowledge. How model APIs actually behave: tokens, caching, tool-use loops, streaming deltas, sub-agents; how each vendor meters and bills; what a cache miss costs. You will know this better than the vendors' own docs.
• Real-time decisions on the hot path. Policy and optimisation applied as the request passes, under a strict do-no-harm rule, with a hard budget on latency and memory and zero tolerance for breaking a developer's session.
• Endpoints and MDM. Certificates, DNS and trust on macOS, Windows and Linux, WSL and containers, distributed through Jamf, Intune, Kandji and friends, coexisting with corporate proxies and zero-trust clients.
• Enterprise operations. It runs in the customer's cloud, not ours: it has to install cleanly, survive without us, fail open when it can't serve, span zones, bound every buffer against hostile input, ship as signed and attested releases, and produce the evidence a SOC 2 auditor asks for.
• Performance. Streaming and back-pressure, bounded decompression, cluster workers, tail latency and capacity at the scale of thousands of developers per customer.
• The console side. A data-plane feature is done when its settings, pages and numbers exist in the product: about a quarter of the role, in Next.js and Postgres.
• The domain is yours. You'll decide what the traffic means, and defend it to the customer's engineers and security team.

02 Why it's interesting

• Nobody has done this before, and the clients change every week; the problem never goes stale.
• The measure is unmistakable: what you ship is scored on real traffic within days, on numbers that appear on a customer's bill.
• Systems problems at every layer, from TLS and binary framing to multi-zone failover to a device's trust store.
• A small, senior, agent-native team, working directly with the CTO, with a direct line to the founders.

03 What you'll do

• Keep us reading every client as their protocols evolve: identity, usage, tools, sessions.
• Own the pipeline that turns live traffic into structured, priced facts, and the taxonomy behind it.
• Build the real-time decision logic on the hot path and the rules that make it safe.
• Make it production-grade: bounded memory, more than one zone, fail-open, crash hardening, hostile-input corpora, end-to-end tests on real endpoints across macOS, Windows and Linux.
• Own endpoint onboarding: installers, trust and steering per OS, MDM distribution, device health reporting.
• Deploy into someone else's cloud and upgrade without drama; produce the operational evidence enterprise security reviews expect.
• Finish each feature in the console: the Next.js and Postgres side of your own work.

04 What we're looking for

• You've built and run backend or infrastructure systems in production, owned them end to end, and care about latency, correctness and observability.
• You're at home on the hot path. Concurrency, streaming, tail latency and failure modes are things you reason about by default.
• You've worked at the protocol level: TLS, HTTP/2, WebSocket, binary formats. Taking an undocumented client apart with a packet capture doesn't scare you.
• You're strong in Node.js and TypeScript, which is our stack, or fluent enough in a systems language that picking it up is quick; and you can ship a page in a Next.js and Postgres product when your feature needs one.
• You understand, or want to go deep on, how model APIs behave and how AI vendors bill: tokens, caching, tools, streaming, credits.
• You've run software in other people's environments: packaging, upgrades, support, the discipline that comes with it.
• You build with coding agents as your normal way of working, and you review their output as carefully as your own.



05 Nice to have

• Proxies, gateways, service meshes or high-throughput streaming systems.
• Endpoint networking and security: PKI, DNS, MDM, OS trust stores, corporate proxies.
• Terraform and AWS, and products deployed in customers' own clouds; SOC 2 controls.
• LLM inference, model APIs or ML infrastructure.
• You've been early at a company before and liked it.