ATPL

AI · Build with AI

Put language models to work inside your product.

For product owners and operations heads who want drafting, summarising and answering done by software inside the tools staff already use, with a person checking before anything leaves.

Describe the problemAll AI capabilities

Rough to structured

Prompt to structure

Generation is the product. Watch a rough ask become fields you can ship — then decide if the rest of the page is worth reading.

What a client typed

We sell about four thousand sarees and dress materials online and every new lot needs a title, a description in English and Hindi, and the size and fabric filled in, which two people do from the supplier’s WhatsApp photos and notes.

Brief ready

What it becomes

Problem
Product titles, bilingual descriptions and attributes written by hand for every new lot.
Scale
Roughly four thousand live products, with new lots arriving weekly.
Systems
The storefront catalogue, supplier WhatsApp messages and photos.
First step
A draft generator inside the catalogue editor; a person approves each listing before it goes live.

Blank page

Paying the blank page tax

Language models drafting, summarising and answering inside your own product.

  • 01 · Draft

    Staff write the same product descriptions, replies and summaries by hand, hundreds of times a month.

  • 02 · Tone

    You tried a public chatbot; it was impressive in a demo and useless without your data.

  • 03 · Cost

    You worry about cost per call, wrong answers, and customer data leaving the building.

A feature, not a chat

Generation inside the product

  • 01

    A feature, not a chat window

    Generation built into the screen where the work happens: the catalogue editor, the support ticket, the quotation form.

  • 02

    Prompt and context design

    What the model sees on each call is fixed in code: your product data, your tone rules, your examples, nothing else.

  • 03

    Output checks

    Every result is validated against a schema and your rules before it is shown; anything that fails goes to a person.

  • 04

    Cost caps and caching

    Repeated context is cached, cheaper models handle the easy cases, and a daily spend cap switches the feature off rather than overspending.

  • 05

    Evaluation set

    Fifty real inputs with agreed good outputs, run before every prompt or model change so quality is measured, not felt.

Generate or write

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.

  1. Fits · 01

    Text your team writes repeatedly from the same kind of input: listings, replies, summaries, quotations.

  2. Fits · 02

    Products where a draft reviewed by a person is useful even when it is not perfect.

  3. Fits · 03

    Teams that can show fifty real examples of what good looks like.

  4. Skip · 01

    Anything that must be exactly right every time with nobody checking, such as a legal notice or a price.

  5. Skip · 02

    Work where the input is a feeling in someone’s head rather than data the model can be shown.

  6. Skip · 03

    Low volumes. Twenty descriptions a month are cheaper to write than to automate.

Your data

Your product data and your examples stay in your database; the model sees only what each call needs. Calls go to the provider you choose under its business terms, which exclude training on your data. We store inputs, outputs and staff edits so quality can be measured, and you can delete them at any time.

How it ships

Prompt, check, ship

  1. gen/01

    Understand

    We collect fifty real examples of the text your team writes today and mark which ones were good and why.

  2. gen/02

    Design

    We write the prompt, the context it needs and the output schema on paper, and agree what the model must never say; the model is chosen last.

  3. gen/03

    Build

    The feature ships behind a review step: staff see the draft, edit it, and their edits feed the evaluation set.

  4. gen/04

    Stay

    We watch cost per call and rejection rate monthly, and re-run the evaluation set whenever a model version changes.

How we work

The brief assistant on this website’s Contact page is built on the same methods: a fixed knowledge block, a schema for the output, caps on cost and length. ATPL will confirm which client work may be named.

    Chosen for the job

    Node.js · Next.js · Python

    Decide the shape

    Before you start

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

    Generative AI for business means a language model producing text or structured data inside a company’s own software: product listings, customer replies, summaries of long documents, first drafts of quotations. ATPL builds it as a feature with a review step, not as a separate chat window.

    Adding one generative feature to an existing product is typically a four to six week build, plus a running cost per model call that depends on the model and the length of the input. ATPL caches repeated context and caps daily spend so the monthly bill is predictable.

    Yes, with testing. Current language models handle Hindi, Gujarati and Marathi usably, but quality varies by language, so ATPL scores each language against real examples before launch and has a native reader review the outputs during the evaluation phase rather than assuming parity with English.

    No. ATPL calls models under business API terms that exclude training on customer data, and sends each call only the context it needs. Your data stays in your database; prompts and outputs are logged in your infrastructure for quality checks and can be deleted on request.

    Continue to