Senior AI Engineer

Arrows; Greater London; On-site; GBP 110,000 - 170,000; Full time

Job summary

Arrows is building a production-grade AI platform for live sports entertainment. You’ll shape architecture, govern AI systems, and drive from experiments to scale.

You will design agent frameworks, data flows, and monitoring with direct access to senior leadership to influence next steps. Reporting to the CTO, you’ll own the end-to-end AI stack, balancing cost, latency, and safety while enabling teams to ship agents quickly and safely, touching data, production deployments, and governance early

Qualifications

• Hands-on experience with agent frameworks in production environments.
• Proven ability to ship AI using LangGraph, LlamaIndex, Semantic Kernel or ADK.
• Strong Python experience on engineering foundations including testing and CI/CD.

Responsibilities

• Lead development of live agents and production-grade AI systems.
• Define governance, cost, latency, and data safety for AI systems.
• Architect and guide the agent estate from retrieval to production.
• Collaborate with CTO and wider teams to deploy AI at scale.

Skills

LangGraph

LlamaIndex

Semantic Kernel

ADK

RAG

Python

CI/CD

Azure

GCP

AWS

SQL

Job description

We are partnering with a profitable, fast-scaling startup shaping the next generation of sports entertainment, recognised by EGR as one of the most innovative startups in gaming, and growing over 15x in the last year.

This is a founding level hire. Reporting directly to the CTO, you'll take the company's use of AI from experiments to a real, production grade capability, shaping the platform and architecture, and setting the governance and standards that keep it safe as the business scales.

What they've built

A product users genuinely love: unlimited group chats, multi game bet builders, and an experience designed around how people actually want to engage with sport, together. The comparison the founders draw is Revolut disrupting Barclays or Robinhood disrupting Etrade; this team is executing the same playbook against the traditional gaming sector.

What you'd be building

Agents that run in a live product, not a slide deck, taken from experiment to something the team is happy to put in front of players, tested and validated to the standard a live product demands. The shared frameworks, templates and infrastructure that let other engineers ship their own agents without starting from scratch. The technical shape of the entire agent estate: how it retrieves, how it evaluates, how it constrains behaviour, how it's monitored, and how any of it reaches production. The sign off process for what goes live, how it handles player data, and where its authority stops.

How you'd work

Sitting with the teams who feel the pain, finding where an agent would genuinely pay for itself, and getting something in front of them fast enough to learn whether you were right. Choosing the approach, low code, pro code, or off the shelf, weighed on cost, control and speed to land. Designing for the failure cases: a call that times out, a tool that errors, a decision the agent should hand back to a person. Acting as the company's AI champion, running office hours, demos and short training sessions that make the wider team better at using AI. With direct access to senior leadership and real influence over where AI goes next, not a backlog someone else wrote.

What they need

Demonstrable impact from agentic or LLM powered systems you've shipped to real users, with the ability to explain what broke and what you changed. Hands on experience with an agent framework such as LangGraph, LlamaIndex, Semantic Kernel or ADK, and RAG in production, including embedding models, vector stores, re ranking, and knowing when a live query beats retrieval. Strong Python experience on a real engineering foundation: testing, version control, CI/CD, and the APIs that serve your own work. Hands on experience with a major cloud and its managed AI services, Azure and AI Foundry or the GCP/AWS equivalents, plus solid SQL and relational modelling. Architectural judgement, making the design call, defending the trade offs, and knowing where an LLM system needs optimising on cost, latency and output that only sounds right. Strong product sense, data driven thinking, and an understanding of what players actually need.

Nice to have

Proving an AI system behaves rather than trusting it: eval sets, output scoring, tracing, regression gates. Retrieval pipelines at volume: embedding at scale, index freshness, accuracy as underlying data moves. Experience with workflows that survive contact with reality: timeouts, failed APIs, a human approving a step.

For more information: [e-mail]