Foundations
Custom AI models: a model fitted to one job in your business
Most of the value of AI models in a business is unglamorous: reading an order from a WhatsApp message, pulling the fields off an invoice, sorting incoming requests, drafting a quotation in the customer’s language, recognising a product from a photo. We fit a model to that one job, measure it, and put it inside the workflow where the job happens.
- Fitted to one job
- Evaluated on your examples
- Private deployment when needed
In short
A custom AI model solution is a language or vision model applied to a single, defined task in your business, with prompts, retrieval, rules and sometimes fine-tuning designed for that task, and an evaluation set that proves it works on your data. It runs inside your workflow, so the extracted order, the classified request or the drafted reply lands where a person would have typed it, for a person to check. It is the engine inside our agents.
What we build
Extraction: orders from WhatsApp messages and photos of lists, fields from invoices, purchase orders and delivery notes, into structured records your system can use. Classification: incoming requests, complaints, leads and tickets sorted by type, urgency and owner. Recognition: a product matched from a photo against your catalogue.
Drafting and summarising: quotations, follow-up messages, payment reminders and replies written in the customer’s language from your templates and data; long threads and documents summarised for a manager. Each is built around the real examples from your business.
- Order and invoice extraction into structured records
- Request, lead and ticket classification
- Product recognition from photos
- Quotation, reminder and reply drafting
- English, Hindi and Gujarati
How we build it
We start with a hundred real examples, not a demo. Those become the evaluation set that every prompt, retrieval change or fine-tune is measured against, so “better” has a number. Confidence thresholds decide what passes straight through and what a person must check.
Models are chosen per task: a large hosted model where quality matters most, a small or fine-tuned one where volume, cost or privacy matter more, deployed privately when your data cannot leave your environment. Prompts are versioned and every change can be rolled back.
- An evaluation set from real cases
- Confidence gates and human review
- Model choice per task, hosted or private
- Prompt versioning and rollback
- Cost and latency tracked per task
What it connects to
The channel the input arrives on, the system the output belongs in, and the review screens where a person checks what the model produced.
- WhatsApp, email, scanned documents and forms
- CRM, ERP, accounting and payroll
- Review queues inside your system
Where it shows up
Automation playbooks this service runs in
Custom AI Models
Common questions
Do we need to fine-tune a model?
Usually not at first. Good prompts, retrieval and rules get most tasks to a usable accuracy. Fine-tuning earns its cost when volume is high, the task is narrow and the examples are plentiful; we will tell you which case you are in.
How accurate is it?
We measure it on your examples and tell you, task by task, before you rely on it. Anything below the threshold goes to a person. Accuracy is tracked after launch too, because your inputs change.
Can it run on our own servers?
Yes. Smaller open models can run privately on your infrastructure or a dedicated server, which is the right choice when the data must not leave your environment. We design for the constraint you have.
Related