Applied AI Engineer
Take an ambiguous idea and turn it into a production system — LLMs, agents, retrieval, media and speech generation, and the backend that keeps it standing up.
The role
Ship the thing. Then make it reliable.
MediaStack puts communication on autopilot — planning, creating, publishing, and optimising it end to end. That means real AI systems carrying real traffic: retrieval over large archives, agents that complete work without a human in the loop, and generation pipelines across text, audio, image, and video.
You will own features from a one-line brief to something in production with users on it. Nobody will hand you a spec. You decide the architecture, build the backend around the model, deploy it, watch it fail, and make it stop failing.
What you will build
Retrieval systems over proprietary media archives — embeddings, vector search, hybrid retrieval, reranking, and the evaluation to prove they answer correctly.
Agents and automation that replace manual editorial and operational workflows, with tool calling, structured output, and sane behaviour when a tool or model misbehaves.
Media and speech pipelines — TTS, audio, image, and video generation wired into queues, workers, and storage that survive volume.
The unglamorous half: APIs, webhooks, Postgres, Redis, background jobs, Docker, deploys, logs, retries, fallbacks, cost and latency budgets.
What we look for
- You have shipped an AI system that people other than you actually used, and you can tell us what it did, who used it, and what broke.
- Strong backend fundamentals — Python, a web framework, SQL, queues, caching, deployment. The model is a component, not the product.
- Practical LLM depth: RAG, embeddings, tool calling, structured output, context management, hallucination mitigation, evaluation. Depth in one real system beats a list of frameworks.
- Production instincts: retries, fallbacks between providers, monitoring, rate limits, concurrency, latency and cost.
- Evidence you learn on your own — side projects, open source, writing, things you built because you wanted them to exist.
Roughly two to five years of relevant engineering. We will not count them if the work speaks louder. TypeScript/Node.js is useful, not required.
This is not
- A research role. We are not publishing papers.
- A notebook role. Prototypes that never left a laptop do not count for much here.
- A prompt-writing role. Prompts are one line in a much larger system.
- An API-wrapper role. Calling a hosted model is the easy part; everything around it is the job.
How we screen
We read your CV against one question: has this person personally built something technically difficult that worked outside their own laptop?
A technical conversation about a system you built — your decisions, your tradeoffs, your failures and what you did about them.
A short, paid, practical exercise close to the real work. No puzzles.
Write your application in specifics. “Built and deployed a retrieval assistant used by 500 people” tells us more than any skills list, and numbers — users, volume, latency, cost — travel further than adjectives.
Apply
Send us the work. Links to what you have shipped matter more than a formatted CV.
Not this role, but still want in? Apply to the group and tell us what you build.