01
Define the task and baseline
Choose one repeated job, the required output, and the errors your team cannot accept. Record current quality, end-to-end latency, review effort, and cost per accepted result.
Enterprise model customization
Fine-tune an open-weight model for your company's work.
I can help your company train a specialist model on approved proprietary internal data, connect it to your workflow, and measure whether it performs better than a frontier model on a specific task.
The goal is better task performance at lower cost, with faster responses. We test those outcomes against your actual workload before committing to a production approach.
By James Perkins. Philadelphia and remote engagements.
Updated .
Specialists inside the company
Frontier models are strong generalists. I expect more companies to use smaller specialist models for repeated work with clear evaluation criteria. A specialist can run by itself, alongside other specialists, or as one step in a broader agent workflow.
Useful candidates include support-ticket classification, document field extraction, internal request routing, and drafting in an approved format. The model handles the defined task, and unfamiliar or uncertain cases follow an explicit review path.
- Step 1Approved input
- Step 2Specialist model
- Step 3Output checks
- Step 4Accept or escalate
Escalation to an external model happens only when your data policy permits it. Otherwise, the workflow stays within the approved boundary and routes to a person.
From an internal dataset to a working service
02
Prepare the internal examples
Review rights and access before using proprietary data. Remove unnecessary sensitive fields, agree labeling rules, and reserve a separate test set that training and prompt iteration never see.
03
Train and compare candidates
Start with a suitable open-weight model and consider LoRA or QLoRA adapters where supported. Compare a tuned model, retrieval, and a strong frontier baseline using the same task and information.
04
Integrate and operate
Connect the approved model to the existing workflow. Add output validation, access controls, exception handling, monitoring, versioned evaluation, and a rollback path before broader use.
The engagement can produce a data-readiness review, a model comparison, a versioned adapter or model where applicable, an integration plan, and an operating runbook. Scope and deliverables are agreed around the task.
Model selection
Qwen3.5-9B and other open-weight candidates
Qwen3.5-9B is one candidate I can evaluate for a compact specialist. The shortlist also includes models from the Gemma, Ministral, Granite, Phi, and gpt-oss families. The best choice depends on the task, license, available memory, context length, and serving environment.
Open weights make local adaptation and deployment possible under the model's terms. They do not make every model equally suitable or remove operating costs.
Questions before a fine-tuning project
Can a smaller model outperform a frontier model?
Yes, on a defined task it can. I compare the tuned model with a strong frontier baseline on examples withheld from training, using the same access to relevant information. Better performance is a result to demonstrate, not a promise for every model or company.
Should our company use fine-tuning or retrieval?
Fine-tuning changes model behavior, such as classifications, extraction formats, or response style. Retrieval supplies current, permission-controlled information at request time. Some workflows need both. A simpler classifier or a prompted model may already meet the requirement.
Does private deployment guarantee that our data cannot leak?
No. Private deployment gives you more control over where processing happens, but training data can be memorized. I help define data access, approved infrastructure, retention, evaluation for disclosure, and controls for logs, model weights, and adapters. Sensitive data is reviewed before any training or transfer.
Will fine-tuning reduce our total cost and latency?
It can, particularly for repeated tasks at sufficient volume. The comparison includes preparation, training, hosting, idle capacity, monitoring, retries, human review, and retraining. I measure end-to-end response times under realistic load. A low-volume workflow may cost less through a hosted model.
Can specialist models work together or by themselves?
Both are possible. A single specialist can handle a validated task. Multiple specialists can classify, extract, and check a result in sequence, with a person or an approved frontier model handling exceptions. Extra calls add cost and latency, so the whole workflow must earn its place.
Bring one task and its current limits.
Please describe what starts the work, the output you need, approximate volume, and the errors or delays that matter most. A high-level description is enough for the first conversation. Please do not send confidential records through the public contact form.
Talk with James about the taskAutomation implementationAgent workflowsDocument processingHow I compare specialists with frontier models
Copyright © 2026 James Perkins. Original page content requires written permission for reuse except where permitted by law. Model names and linked materials belong to their respective owners. Content rights and permissions.