We Are Potato

AI

AI in development, and in what's already live

Some days a team loses hours hunting a number, a contract, or the report nobody had time to build. AI comes into development to take repetitive work off the path. It comes into the product when the answer, the search or the report has to work inside the system that's already live.

In production

What we've already put to work

01Chat over the data

Ask in plain words, answer from the database

The people deciding shouldn't depend on a query only one person can write. The chat takes the question, reads the operation's data and answers with the number and where it came from.

  • Free-form questions, no filters to learn
  • The query shows next to the answer
  • Anything out of pattern comes flagged

Finance panel · chat

What was the margin per product line in September, and what fell out of pattern?

  • Reading monthly_sales
  • Joining with line_costs
  • Comparing with the previous 6 months

September's margin closed at 31.4%. One line came in below where it usually sits.

LineRevenueMargin
Kits$82k36.2%
Singles$54k33.9%
Wholesale$37k22.1%

Wholesale: 8 points under the average of the last 6 months.

Ask about revenue, margin, stock

Made-up screens and data. The shape is that of systems we've already built.

02RAG and embeddings

An answer that points at the document

Contracts, manuals and policies exist, but nobody finds them. Documents become passages indexed by meaning, search brings back the closest ones and the answer cites where it came from.

  • Search by meaning, not only by keyword
  • Every answer carries file and page
  • No passage found, and the answer says so

Document base · search

How long do we have to return a defective batch?

Passages found

  1. 1

    supplier-contract.pdfp. 794%

    Defective batches may be returned within 15 calendar days of receipt.

  2. 2

    quality-policy.pdfp. 381%

    A return requires a report from the quality team attached to the order.

  3. 3

    logistics-manual.pdfp. 1267%

    Return freight is the supplier's cost when the defect is confirmed.

Answer

You have 15 calendar days from receipt 1. The return needs the quality report 2, and freight is on the supplier if the defect is confirmed 3.

Made-up screens and data. The shape is that of systems we've already built.

03Document reading

From PDF to fields the system understands

Invoices, contracts and reports arrive as files. The AI reads, picks out what matters and returns fields ready to check. The person running the process confirms, and the system stores.

  • Amount, date and clause marked on the original
  • A doubtful field waits for review
  • Confirming is the operator's call

invoice-48213.pdf

Supplier: Serra Azul Metalworks

Issued: 12/09/2026

Total amount: $3,690.00

Due: 12/10/2026

Terms: 30 days, 2% penalty after due date

Fields read

Issued
2026-09-12
Amount
3690.00
Due
2026-10-12
Penalty
2% / month

Penalty: the text doesn't say if it's monthly. Waiting for a check.

Made-up screens and data. The shape is that of systems we've already built.

04Automated pipeline

The report goes out and nobody has to remember

A flow runs on the agreed rhythm: pulls the data, compares it with history, flags what fell out of pattern and sends the written report. Runs are logged like any other service.

  • Scheduled, with a run history
  • The anomaly comes with the number behind it
  • If it fails, it tells someone

Flow · weekly report

  1. Pulls the dataSales, stock and costs, every Monday at 6am
  2. Compares with historyAverage and variation over the last 12 weeks
  3. Flags what fell out of patternOnly what crosses the agreed limit
  4. Writes and sendsA short text for whoever decides, with the numbers

Latest runs

  • Mon 28/09 · 06:02Completed, 2 alerts
  • Mon 21/09 · 06:01Completed, no alerts
  • Mon 14/09 · 06:04Failed reading stock, team notified

This week's report

Revenue is up 6% over the 12-week average.

Wholesale sold 11% less, and the label stock is the likeliest cause.

Two items below the reorder point.

Made-up screens and data. The shape is that of systems we've already built.

05Structured output and tools

The AI returns fields, not paragraphs

To connect AI to a system, the result has to fit in a table. The output has a defined shape and is validated before it's stored. When the AI calls a tool, it gets the permission the task needs and nothing more.

  • Shape defined by a schema
  • Validation before storing
  • Tools with limited permission

Order intake · email

Email received

Hi, I need 40 boxes of the large kit by Friday, delivery to Contagem. Please invoice Casa Lume.

What the AI returns

{  "product": "large-kit",  "quantity": 40,  "unit": "box",  "delivery": "2026-10-09",  "city": "Contagem",  "invoice_to": "Casa Lume"}

Validation

  • okProduct exists in the catalog
  • okQuantity is a whole number
  • reviewDelivery in 4 days, under the standard 7

Order draft created. Lead time sent to sales for a check.

Made-up screens and data. The shape is that of systems we've already built.
  1. 01

    Starts with the problem

    Before the AI comes the question: which part of the process deserves the time saved? Sometimes the answer is a plain rule, no model at all.

  2. 02

    Answers with a source

    Where it can, the answer shows where it came from: the query, the file, the page. Whoever reads it can check.

  3. 03

    Tested on real cases

    Real questions, real documents and the cases that tend to go wrong are in automated tests before delivery.

  4. 04

    The operator decides

    The AI proposes, the operator confirms. The system records what was suggested and what was accepted.

  1. 01

    Do you use AI?

    Yes. Claude, Codex, Cursor, Lovable and Figma Make come in at different stages of the work, depending on the project.

  2. 02

    Does it change price and timeline?

    It can. Some tasks that used to take longer can now be done faster, but that depends on the kind of work and on how much these tools help in that case.

  3. 03

    What about quality?

    Anything done with AI support also goes through review and testing before delivery.

  4. 04

    Can you put AI into my project?

    In some cases, yes. It might be a chat over the data, document search, file reading or a report that goes out on its own. First we understand the need and assess whether AI makes sense there.

Day to day
  • Claude
  • Codex
  • Cursor
  • Lovable
  • Figma Make
  • MCP
In the applications
  • GPT
  • Claude
  • OpenRouter
  • RAG

Got an idea involving AI?

Tell us