Hire an LLM fine-tuning engineer: DAPT, SFT and DPO
I trained commerce foundation models at Shopify and fine-tuned the LLMs behind Sidekick, Shopify's AI assistant for merchants, from data curation through domain-adaptive pre-training, SFT and DPO. I help AI-native teams make an open model fluent in their domain, and prove it with evaluations.
A good fit if…
- A general model doesn't know your domain's language, products or formats, and prompting has hit its limit.
- You want a smaller, cheaper model that matches a frontier model on your task.
- You have data, but not the experience to curate it into a training set that works.
Look elsewhere if…
- Your problem is the model not finding the right information. That's retrieval, and fine-tuning won't fix it.
- You have no evaluation yet. You'll need one first, and I can help build it.
- You need GPU infrastructure managed. I size and use compute, but I'm not an infrastructure provider.
Deliverables your team keeps.
From brief to handover.
Define the target tasks and build the evaluation. Measure how far prompting and retrieval get you.
Curate data and run small training experiments to find what moves the metrics.
Full training runs, evaluation against the baseline, and iteration.
Model, data recipe, training scripts and evaluation harness in your repositories.
Works with what you run.
What I've built before.
How it's priced.
Not sure where to start? Most teams begin with a fixed-scope search relevance audit.
Starting prices; the final quote depends on scope and follows a free intro call.
Remote, working with teams in the US, UK and Europe. Async-first, with live time during your Pacific mornings, or your morning and early afternoon if you're in Europe. Contracted and invoiced in USD through Infersense Labs LLC (Wyoming, USA).
Questions buyers ask.
Fine-tuning or RAG?
Often both, for different jobs. RAG gives a model fresh knowledge it can cite; fine-tuning shapes its language, format and behaviour. I decide with evaluations, not opinions.
Do we need our own GPUs?
You'll need compute for training. I help you size it and choose between your cloud, managed fine-tuning services and rented GPUs.
How much data do we need?
It depends on the method. SFT can work with thousands of good examples; DAPT needs far more domain text. The first weeks answer this for your case.
Who owns the model?
You do. Weights, data recipes and scripts live in your accounts and repositories.
Tell me what you're hiring for.
A free 30-minute call with me. You'll leave with a clear next step, and an honest answer on whether I'm the right person for it.