MLOps &
Fine-tuning
I build the pipeline around a model: fine-tuning it on your data, measuring whether it actually improved, and shipping it reproducibly — so a better model is a repeatable process, not a one-off experiment.
From dataset to deployed model, reproducibly
Fine-tuning is only useful if you can prove it helped and ship it reliably. I run QLoRA fine-tunes on your own hardware, build the dataset and evaluation harness that tells you whether the new model beats the old one, and package the result with a reproducible merge-then-quantize deployment.
The whole loop is designed to be run again: new data in, evaluated candidate out, promoted only if it wins. No guessing, no silent regressions.
What's included
Stack
PyTorch · Transformers · TRL · Unsloth · PEFT / QLoRA · datasets · llama.cpp (quantize) · Python · CUDA
Frequently Asked Questions
When is fine-tuning worth it?
When prompting and retrieval have plateaued and you need consistent behaviour, a specific format, or domain knowledge the base model lacks. I'll tell you honestly if RAG or a better prompt would get you there more cheaply first.
Can you fine-tune on-prem?
Yes. QLoRA makes it feasible to fine-tune capable models on a single workstation GPU, entirely within your environment — your training data never leaves.
How do you know the new model is better?
A held-out evaluation harness with task-specific benchmarks, compared head-to-head against the current model. A candidate is only promoted if it measurably wins.
What do I get at the end?
A deployable quantized model plus the pipeline to reproduce it — dataset, training config, eval results and the merge-then-quantize steps.
