A real toolbox. Frontier model APIs and open-weight models, GPU budget for training and serving, plus internal agents, MCP servers and evaluation tooling built in-house against our own data. Time to evaluate new tools — we adopt fast and drop fast.

You set the practice. This is a new role at Tango. How we evaluate models, gate releases, version prompts and datasets, and decide build-vs-buy is not written yet. You write it, and you share it with the wider engineering org. No ceremony. A feature starts as a goal, not a ticket. You decide with the product manager what is worth building, and then you build it.

Responsibilities

Own the problem, not just the model. With the product manager, turn a goal into a solution: question the proposed approach, offer alternatives, and say what to leave out.

Build with LLMs. Context engineering, retrieval, tool use and agentic flows, structured outputs, guardrails — plus fine tuning and distillation when a general model is too slow, too expensive, or not good enough.

Train your own models. When a small, narrow, purpose-built model wins — moderation and safety classifiers, ranking and recommendation, audio and vision, abuse and anomaly detection — you own it from data collection and labelling strategy through training and evaluation.

Decide build vs. buy, with numbers. Frontier API, open-weight model, or trained in-house: justified on quality, latency, unit cost and data-privacy constraints, not on preference.

Ship it to production. Serving under real load, latency and cost budgets, batching and quantization, safe rollout, and integration with the backend services and clients the feature touches.

Prove it works. Offline evaluation sets you build and defend, then online A/B. Monitoring for drift, regressions and feedback loops once it is live, and the honesty to roll something back.

You will not do this alone, and you are not expected to be an expert in all of it. Domain experts in backend, platform, infrastructure, data, web, Android and iOS are there to be asked, and those teams maintain the tooling, conventions and test platform for each surface.

Requirements

What we're looking for

LLM ENGINEERING

• Hands-on production experience with LLMs — not demos: retrieval, tool use and agents, structured output, prompt and context engineering under latency and cost constraints
• Practical adaptation of models: SFT / LoRA / preference tuning, and distilling a large model into a small one you can actually afford to serve
• You build evaluations before you build confidence: task-specific datasets, sensible use of LLM-as-judge and awareness of where it lies, offline-to-online correlation
• Clear-eyed about failure modes — hallucination, prompt injection, silent quality drift, benchmark overfitting — and you verify rather than trust

TRAINING YOUR OWN MODELS

• You have trained and shipped models on real, messy production data — classification, ranking or recommendation, CV, speech/audio, or abuse and anomaly detection
• Full ownership of the data path: collection, labelling strategy and label quality, splits, leakage, class imbalance, and the cost of getting these wrong
• Solid ML fundamentals and the ability to debug a model rather than retrain it and hope
• Fluent in Python and PyTorch, and efficient with GPUs — you know what your training run costs and why

PRODUCTION ENGINEERING

• 5+ years of professional engineering overall, with genuine software engineering skill — this role does not end in a notebook
• Serving models under real load: throughput, tail latency, batching, caching, quantization, and the trade-offs between them
• Reproducible pipelines and orchestration; versioning of data, models and prompts; the ability to explain what exactly was deployed last Tuesday
• Production monitoring for ML: drift, regression detection, feedback loops, and safe rollback
• Comfortable working with the services and clients around your model — backend, web, Android and iOS — with AI tooling and expert support

English proficiency at Intermediate level or higher

HOW YOU WORK

• Daily, hands-on use of AI coding agents in real production work, and the ability to scope tasks well for them Comfortable owning a feature end to end: design, implementation, evaluation, rollout, and the result You have lived with what you shipped rather than handing it over at the end
• Direct, easy collaboration with product managers, designers and data analysts

Would be great to have

• Real-time or streaming inference on live video and audio; on-device or edge models
• Content moderation, trust & safety, or anti-fraud at consumer scale
• Recommendation and ranking systems for consumer feeds, with online experimentation
• Multimodal work: video, audio, speech, or text-and-image together
• Serving and training infrastructure: vLLM, Triton, TensorRT, Ray, Kubeflow or similar
• Building or improving agentic development workflows — custom agents, MCP servers, evals — and sharing what works with other engineers
• Enough Java 17+ / Spring Boot to integrate cleanly with our backend, or readiness to ramp up quickly Live-streaming, video, high-scale consumer products, fintech or communication platforms
• Docker and Kubernetes, GCP and BigQuery, Aerospike

Technical stack

ML: Python, PyTorch, Hugging Face, GPU training and inference on GCP, experiment tracking and model registry, BigQuery for data.

Platform: Java 17–21 and Spring Boot, GCP, Apache Kafka, MySQL, Redis, Aerospike, Docker, Kubernetes, GitLab. Clients: React and TypeScript on web, Kotlin and Jetpack Compose on Android, Swift on iOS.

How we hire

• An intro call.
• A 90-minute working session on a small, real problem over live-streaming data, which we hand you at the start. You work on your own machine, with your own AI tooling, sharing your screen. We are interested in how you frame the problem, what baseline you choose, what you ask us, and how you satisfy yourself that what you built actually works. Fluency with our particular stack is not scored, and there is nothing to prepare.
• A final conversation with the team, including a deep dive into a model or AI feature you took to production.

What we offer

• Stock options — every employee is a partner in the company
• Competitive salary and performance review bonus
• Hybrid working model: one day per week remotely
• Official employment in Cyprus: work permit visa for employees and residence permit for spouses Medical insurance (100% for employees, 75% for family members)
• Comfortable office in Limassol centre with lunch, dinner and snacks
• Opportunities for professional growth, Greek lessons, sport compensation
• Family corporate events — we believe our spouses are part of our professional success
• Tango is Title Sponsor of Aris Limassol F.C.: free attendance at matches, VIP and Sky Box access, events and merchandise
• Gifts for employees