Classify and route requests
Learn your support categories and escalation rules from approved, labeled examples. Check rare categories and incorrect routing before automating handoffs.
PHILADELPHIA / SPECIALIZED MODELS
A model trained for the work your company does.
I can help you fine-tune an open-weight model on approved proprietary internal data for a specific business task. Through Philly Si Guy and Automation Transformation Consulting, I help companies test whether a specialist can deliver better results, lower total cost, and faster responses than their current frontier model workflow.
- 01Approved examples
Representative inputs and reviewed answers.
- 02Adapt and test
Train on one set. Evaluate on unseen cases.
- 03Route the result
Accept within agreed limits. Escalate exceptions.
A specialist can work alone, with other specialists, or with a frontier model and human reviewer.
01 / THE OPPORTUNITY
Give a specialist a clearly defined job
Frontier models are impressive generalists. I expect more companies to use smaller specialist models for work that repeats and can be evaluated. A specialist might classify an incoming request, extract a defined set of fields, or draft an answer in an approved format.
Your examples can teach a model the distinctions, terminology, and output conventions that matter in that task. The objective is a measurable improvement on your work. A higher public leaderboard score alone does not establish that.
There is a real possibility of narrow-task outperformance, including lower inference cost and lower response time. We need to measure all three separately, under the conditions your team would actually use.
02 / POSSIBLE STARTING POINTS
Tasks with a result you can check
Extract consistent fields
Turn recurring business documents into a defined structure. Validate required fields and keep unusual or incomplete documents in a review queue.
Draft in an approved format
Create first drafts of internal summaries or routine responses using examples your team has reviewed. Measure factual mistakes and editing time.
Use internal information appropriately
Combine retrieval with a specialist when the work needs current documents as well as consistent behavior. Preserve document permissions and source references.
These are candidate workflows, not claims about completed client projects. Consequential decisions need separately scoped controls and review.
03 / THE ENGAGEMENT
From company examples to a tested workflow
Define the task and baseline
Agree on acceptable output, error costs, expected volume, and response-time needs. Test the current process, a well-prompted frontier model, and simpler alternatives.
Prepare the approved data
Check data rights, access, sensitive fields, duplicates, and label consistency. Separate training examples from validation and a final held-out test set.
Adapt a suitable model
Compare candidates such as Qwen3.5-9B against your hardware, license, language, and task requirements. Parameter-efficient methods such as LoRA or QLoRA may reduce training resource needs.
Evaluate the whole process
Measure task quality, failure cases, response times, human correction effort, and total operating cost. Test any proposed quantization on the actual serving setup.
Integrate with an owner and fallback
Define permissions, monitoring, versioning, review queues, and rollback. Use an approved frontier model only where the company's data policy permits the call. Otherwise route exceptions to approved local processing or a human reviewer.
04 / PROPRIETARY DATA
Decide where the data and model can go
Open weights give a company more deployment options. They do not automatically provide privacy, unlimited commercial rights, or a system that is inexpensive to operate.
Before training, we should agree on where examples are stored, who can access them, whether a managed provider is involved, and how logs, backups, adapters, and checkpoints will be handled. A trained artifact may still expose sensitive information.
Training data should be relevant and deliberately selected. I would avoid turning an entire internal document collection into training material when a permission-aware retrieval system would better fit the problem.
05 / THE BUSINESS CASE
Compare cost per accepted result
A small model can cost less to serve, but hardware, utilization, data preparation, training, evaluation, maintenance, and human review all count. A lightly used self-hosted model may cost more than a metered API.
For speed, I would measure time to first token and end-to-end completion at your expected concurrency. Extra reasoning, retries, retrieval, or multiple specialists can erase a model-size advantage.
See the comparison and evaluation plan for a practical scorecard, then review the open-weight model shortlist.
METHODS / PRIMARY SOURCES
Research behind the approach
A 2024 Microsoft Research case study found that small fine-tuned models could beat larger models on accuracy and running time in an internal cloud supply-chain application. It demonstrates the pattern in a historical task, not a comparison with today's strongest frontier models or a result from my client work.
LoRA and QLoRA describe adaptation methods that use fewer training resources. Retrieval-augmented generation combines generation with an external knowledge source. None of these sources establishes an outcome for your company.
- Microsoft Research: Small Language Models for Application Interactions: A Case Study (2024)
- LoRA: Low-Rank Adaptation of Large Language Models
- QLoRA: Efficient Finetuning of Quantized LLMs
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Sources reviewed October 12, 2026. Model and license details are linked on the shortlist.
QUESTIONS / ANSWERS
Before we get started
Can a small fine-tuned model outperform a frontier model?
It can on a narrowly defined task with suitable training examples. I would compare it with a well-configured frontier baseline on held-out company cases. Better task quality, lower total cost, and faster responses are goals to test, not guaranteed outcomes.
Can you fine-tune a model on our proprietary internal data?
I can help assess the data, prepare approved examples, adapt an appropriate open-weight model, and evaluate the result. The project starts with rights, access, retention, hosting, and a defined business task. Please do not send proprietary data through the public contact form.
Does private hosting make a model completely private?
No. Privacy depends on the complete system: data access, logs, backups, hosting, support access, model artifacts, and any external services. Fine-tuning may also make sensitive examples recoverable, so data minimization and access controls remain necessary.
Do we need fine-tuning or retrieval-augmented generation?
Retrieval is often useful when answers depend on changing documents and need source references. Fine-tuning is a candidate for repeatable behavior, terminology, classification, or output structure. A project may combine them, or use a simpler method after evaluation.
Do you work with Philadelphia companies and remote teams?
I am based in Philadelphia. I can discuss a focused model and automation project with local organizations or remote teams, with delivery and collaboration agreed during scoping.
CONTINUE / LEARNING
Plan the next step
PHILADELPHIA / WORK WITH ME
Describe one task you want to improve
Tell me what starts the work, what a good result looks like, roughly how often it happens, and where the current approach falls short.
Please keep proprietary records, customer information, credentials, and internal documents out of the public contact form. We can agree on a suitable data review process after the initial conversation.
Discuss a private model project