Service

LLM Fine-tuning, Inference and Deployment

SFT and LoRA fine-tuning, evaluation harnesses, and serving that fits the budget.

What I build

Smaller, cheaper models that behave like your best prompt. I mine successful trajectories from production logs, curate them, fine-tune with SFT and LoRA, and prove the result on a held-out evaluation set before anything is deployed.

How I work

Evaluation first. Nothing changes until there is a number to compare against. Then I iterate on data quality rather than hyperparameters, because that is where the gains are. Serving is chosen last, once we know how small the model can be.

When to call me

Your LLM bill scales faster than your usage, your agent needs reliable tool calls, or you need a model that runs inside your own infrastructure.

What you get

Gold evaluation dataset and scoring harness
Curated training set mined from production traces
Fine-tuned model (SFT, LoRA or QLoRA) with a before-and-after report
Serving setup sized to your traffic and cost target

Selected work

Case studies for this service

hypREspace

Measured

hypREspace: Cutting Token Costs 90% While Raising Tool-Calling Accuracy to 95%

Dynamic prompt injection, a gold evaluation set and Amazon Nova fine-tuning cut a live agent's token spend by 90% and lifted tool-calling accuracy from 75.8% to 95%.

Measured outcomes

Measured with before-and-after per-query token counts and tool-call accuracy scored against the gold evaluation set.

90%

Token cost reduction, measured per query

75.8% to 95%

Tool-calling accuracy on the gold eval set

Read case study

Related work

Related projects

Fine-Tuned Tool-Calling Agent

A Small Model Taught to Use 40+ Tools

Self-Built ProjectDemonstrated

A LangGraph ReAct agent over a custom MCP backend exposing 40+ tools, with the decision model replaced by a fine-tuned Gemma 3 4B. Successful multi-turn tool trajectories from production were filtered and reformatted into an SFT dataset, then used to train Gemma 3 4B with QLoRA — cutting cost and latency while making tool-call formatting far more reliable.

Gemma 3 4BQLoRASFT

Questions I get asked

When is fine-tuning worth it?

When a prompt-steered model is accurate enough only with a huge prompt, or when you need consistent structured output at lower cost. If a better prompt or retrieval fixes it, I will tell you and stop there.

Which models and tools do you use?

Open models such as Gemma and Llama with Unsloth or Hugging Face TRL, and Amazon Nova on Bedrock for managed fine-tuning. Evaluation runs with programmatic checks and an LLM judge on a held-out set.

How do you measure success?

Token cost per query, task accuracy on the gold set, and latency, all recorded before and after. In one engagement that meant a 90 percent token cost cut and tool-calling accuracy up from 75.8 to 95 percent.

Have a problem that looks like this?

Tell me about it. I reply within one working day with a first take and no sales pitch.

Book a call Send a message