ATPL

AI · Start right

A model that speaks your domain.

For teams whose work has its own vocabulary, formats and standards: textile grades, pharma SKUs, legal clauses, Hinglish support tickets. The model is adapted to them, after prompting alone has been shown to fall short.

Describe the problemAll AI capabilities

Examples to tone

Examples become a voice

The hold-out check and the side-by-side with the base model — tune only when prompts are not enough.

What a client typed

Our support team replies in a mix of Hindi and English with our own product names and policy wording, and the drafts from the general AI sound like a bank; we want replies that sound like us and get the product codes right every time.

Brief ready

What it becomes

Problem
Generated support replies miss the house tone and misuse product codes.
Scale
A support team replying daily; a history of thousands of past replies to learn from.
Systems
The helpdesk or WhatsApp Business inbox, the product catalogue, the policy documents.
First step
Score the best prompt plus catalogue retrieval first; fine-tune only if the held-out test shows it is still short.

Off-the-shelf misses

When the generic model misses your voice

Models tuned to your vocabulary, formats and judgement, when prompting is not enough.

  • 01 · Generic

    The general model is nearly right but keeps mangling your product codes and your way of writing.

  • 02 · Tone

    Your prompts have grown to three pages and still miss the edge cases.

  • 03 · Leak

    Each call is slow and expensive because the model needs so much explaining.

Your voice, your terms

A model tuned on your examples

  • 01

    Prompt-first comparison

    Before any tuning, the best prompt and retrieval setup is built and scored, so you know what customisation adds and whether you need it at all.

  • 02

    Training set from your work

    A few hundred to a few thousand examples drawn from your real outputs, cleaned and labelled with your team, with a held-out test split.

  • 03

    A tuned model

    An open-weight model fine-tuned in your cloud account, or a hosted model tuned through its provider, chosen by privacy need, cost and latency.

  • 04

    Side-by-side evaluation

    The tuned model, the prompted model and your current process scored on the same held-out cases, with the cost per call beside each.

  • 05

    Retraining recipe

    The data pipeline, scripts and evaluation set handed over so the model can be refreshed when your products or formats change.

How it ships

Examples, train, compare

  1. tune/01

    Understand

    We collect examples of the output you consider right and, just as important, the ones you consider wrong, from the people who judge them.

  2. tune/02

    Design

    We decide the target behaviour, the evaluation set and the cheapest route to it; fine-tuning is chosen only after prompting and retrieval are measured.

  3. tune/03

    Build

    Training runs in iterations against the held-out set; the model ships behind the same service layer so it can be swapped back if it regresses.

  4. tune/04

    Stay

    Output is sampled monthly, new examples are added from corrected cases, and the model is retrained when the measure slips or the base model is retired.

Tune or prompt

Where it fits, where it doesn’t

Saying when this is the wrong tool is the fastest way to be believed about when it is the right one.

Fits when

  • A stable, specific style or format the model must reproduce thousands of times.
  • A vocabulary the general model gets wrong: codes, grades, local terms, mixed-language text.
  • Enough examples of right and wrong output to train on and to test on separately.

Doesn’t fit when

  • Knowledge that changes weekly; put that in retrieval, not in the model’s weights.
  • Problems a better prompt fixes. Fine-tuning is the last tool, and we measure before reaching for it.
  • Fewer than a few hundred good examples; the model would learn noise.

Your data

Training examples are drawn from your own records and stay in your cloud account; an open-weight model tuned there never leaves it. If a provider-hosted model is tuned, it is under that provider’s business terms, which exclude using your data to train their general models. The tuned model and the training set are yours.

How we work

ATPL measures prompting and retrieval before it customises a model, and reports when tuning is not needed. ATPL will confirm which tuned models in production may be named.

    Chosen for the job

    Python · Node.js

    Decide the shape

    Before you start

    Plain answers on scope, data and what a first proof usually takes.

    Fine-tuning is training an existing language model further on your own examples so it adopts your vocabulary, format and judgement without being told in every prompt. ATPL uses it after prompting and retrieval have been measured and found short, and tunes in your cloud account or through the provider, depending on privacy and cost.

    Fine-tune when the problem is style, format or vocabulary the model must reproduce consistently, and prompts have grown long and still miss cases. Use RAG when the problem is knowledge that changes. ATPL builds the prompted and retrieval versions first and recommends tuning only when the side-by-side score justifies it.

    A few hundred clean, representative examples can shift style and format; a few thousand are needed for harder judgement tasks. Quality matters more than quantity, so ATPL spends most of the effort cleaning and labelling with your team and holds back a test set the training never sees.

    Yes. Mixed Hindi and English support messages, Gujarati product descriptions and similar text are good candidates, because general models often get the mix wrong. ATPL builds the evaluation set in the same language mix as your real traffic and has native readers score the outputs.

    Continue to