About the role
Inku is hiring an ML Platform Engineer. Someone who can support our AI stack for fine-tuned, self-hosted models running across our platform, and owns everything in between: training, infra, deployment, serving and making them available to every part of the product.
Inku is a brand intelligence platform. Underneath it sits a Brand Engine and a multi-agent creative pipeline that turns a brand's identity, visual grammar and product rules into on-brand output for brand. The data is what makes us good. We want models that are trained on it, run where our customers need them to run, and that we control end to end.
What's in it for you
- Own a greenfield domain: there is no fine-tuning fits all playbook, you write it with us.
- Full ownership of the model stack, from dataset to GPU cluster to production endpoint
- Work on a genuinely hard problem: teaching models aesthetic judgment and brand consistency, not just accuracy
- Ship models that run on our own infrastructure, real deployments, real constraints
- High ownership, no layers, you work directly with the founders in a team.
- Build with the best AI tooling available and be fast because of it
The finetuning work
This is the core of the role. Inku's models need to learn things that don't come out of a base model: what makes an image on-brand for one brand and off-brand for another, how a brand writes, which product rules are hard constraints and which are taste.
- Build the training data. Turn our data and human quality signals into clean, versioned datasets. Decide what to include, what to weight, what to leave out.
- Fine-tune the right model for the job. LLMs for the agents and text surfaces, diffusion and vision models for the creative pipeline. LoRA/QLoRA where it's enough, full fine-tunes where it isn't. You know the difference and can argue it.
- Design evals that mean something. Loss going down is not the goal. You build evaluation sets and judges that measure brand consistency and visual quality, catch regressions before they ship, and tell us whether a new model is actually better.
- Iterate fast. Short training loops, clear experiment tracking, and a feedback path from what brands see in the product back into the next dataset.
The impact you make
You are more than an MLOps engineer. You are the person who makes Inku's own models real.
- Fine-tune and evaluate open-weight models on Inku's proprietary data. LoRA/QLoRA and full fine-tunes, dataset curation, eval harnesses, regression tracking.
- Build and run the serving layer: vLLM / TGI / equivalent on GPU infrastructure, autoscaling, batching, quantisation, latency and cost trade-offs.
- Help decide our GPU strategy: cloud vs. owned, what we train on, what we serve on, and what it costs.
- Make models available across the platform through our model gateway, so agents, the API and the SDK consume them exactly as they consume hosted models. Swap-in, not rewrite.
- Set the standard for how Inku trains, ships and monitors models going forward.
Note that this is not a research role. We care about models that serve production traffic, not notebooks.
What you bring
- Hands-on experience fine-tuning open-weight LLMs and ideally diffusion or vision models, and getting them into production
- You've built training datasets from messy real-world data and designed evals that catch quality regressions, not just improvements in loss
- Solid infra background: Kubernetes, containers, GPU scheduling, cloud (GCP preferred), infrastructure-as-code
- Fluency in the serving ecosystem and the judgment to know when each one is the wrong choice
- Strong Python; you can move through a codebase without a tour guide
- Independent thinking: drop into an unfamiliar codebase and after an hour you're asking the right questions instead of waiting for instructions
- You use AI tools to build and it makes you fast
- Strong communication in English; Dutch is a plus
- Based in Belgium or willing to relocate; we don't hire fully remote
Nice to have: RLHF / DPO or other preference-tuning experience, eval design for generative quality, MLOps tooling like MLflow / Weights & Biases, experience with image-generation pipelines.
Why join Inku
Less talking, more shipping.
- Driving seat: you define how Inku does models. As we grow, this is the foundation of an ML platform team, and you'd be the one who built it.
- Ownership from day one: real decision space, and we expect you to use it.
- Work that ships: every model you train and deploys ends up in front of brands and their customers within weeks, not quarters.
- Competitive package: a competitive salary plus extras like a mobility budget or company car, and the usual benefits on top.