Homepage Smart Data

Smart Data

Headline Ask your data a question. Get the table. And the proof., next to a laptop showing Data Chat answering which stores are behind their quarter target, and a card explaining an order plan

Services:

  • Product design
  • Full-stack development
  • AI engineering (agents, NL-to-SQL)
  • Data engineering
  • DevOps

Industry:

  • Retail chains · Any data-driven business

Technology:

  • Node.js / Express 5
  • PostgreSQL (Cloud SQL)
  • React 19 + TypeScript
  • Google Cloud Run, Firebase
  • OpenAI, Anthropic, Google models, switchable per agent
  • MCP server
  • RTL and LTR interfaces

Challenge

  • “We have dashboards. Every new question still becomes a ticket to the BI developer and a two-week wait.”
  • “An assistant that is wrong 5% of the time is useless to my CFO. It has to be right, or say it cannot answer.”
  • “Every chain’s data is different: 30 million rows, list prices instead of takings, headers that change without warning.”
  • “Adding our next client must mean configuration, not a fork of the product.”

Results

  • Store managers ask in plain language and get a table, an Excel export and the SQL that produced it, under a hard 15-second guard, so questions are answered in the meeting, not two weeks later
  • 30M+ sales rows in one client’s fact table and about 30M records imported nightly on a modest configuration, so the answers cover all the data, not a sample
  • Every report carries a confidence score and its unresolved fact-checks; every money figure states its basis and data-through date, so finance can trust what it reads
  • 7 retail chains run on one codebase and one deployment; ten concurrent conversations with no slowdown in load tests
  • The 2nd and 3rd capability shipped in weeks thanks to the module system; 1,900 commits in eleven months

Good fit for

  • Retail and distribution chains with many stores and a BI tool nobody can question
  • A CFO who has been burned by a confident dashboard that was wrong
  • Buyers who want the purchasing recommendation and its arithmetic on one screen
  • Software companies that need a multi-tenant AI product on one deployment

It is Monday, 10:04. Your regional manager asks which stores are behind their quarter target. In the old world that is a ticket, a BI developer and a two-week wait. Here she types the question and gets a sentence, a table, an Excel export and the SQL that produced it, before the next agenda item. Smart Data is an AI employee that reads your real data, answers in plain language, writes its own investigative reports and, unlike most assistants, says so when the data cannot answer.

The real Data Chat answer with the store table beside a four-step flow from plain question to table, Excel and query; figures 30M+, ~30M, 15 s

Asked at 10:04. Answered, with the table, at 10:04. Data Chat turns a plain question into SQL against the client’s own schema, runs it under a hard 15-second guard and answers with a formatted table, an Excel export and the query it used, in Hebrew or English, right-to-left or left-to-right. Materialized views turn minute-long scans into millisecond reads, so a chain with 30 million sales rows answers as fast as a spreadsheet. About 30 million records arrive every night on a modest server. Point-and-click dashboards sit beside the chat for the people who prefer them, driven by a whitelisted semantic model so the browser never sends SQL.

Dark infographic: capability manifest, deterministic gate, plan-query-write, four independent guards, confidence score; with the real unresolved-fact-check banner and the 40% confidence card

The part competitors skip: an honesty layer, enforced in code. Each dataset carries a capability manifest: which measures are exact, which are estimates and by how much, which dimensions are missing, and the client’s own vocabulary. A deterministic gate refuses impossible questions before any SQL is written. Intelligence reports plan, query, write and then verify themselves through four independent guards, including a separate fact-checking model, and show the reader their confidence score and whatever could not be reconciled. Every money figure states its basis, its data-through date and whether the last day is partial. Invented test cases passed 65 of 65; one real session found two silent wrong answers, which is why every real customer question is frozen into a replay corpus and re-run before release.

Diagram: replenishment engine feeding screen, chat and weekly report, above the real procurement line for Umbrella folding with its 'order 48, rounded up to cartons of 36, 72 units' explanation

One engine. Three surfaces. The same number everywhere. Smart Replenishment computes what to order, how much and by when, nightly, from sales velocity, stock, open orders and the lead times buyers set themselves, with a “Why?” under every line that shows the arithmetic. One engine serves the procurement screen, the chat tool and the weekly report, so the same question gets the same number everywhere. Business users describe a new screen in a sentence and the model returns a validated spec, never raw HTML or SQL; the server compiles the queries, probe-verifies the totals and renders it in the client’s branding. Smart Tune shows the 144 items a change will touch and waits for approval. A personal link even turns the platform into an MCP server, so the client’s team can build screens from ChatGPT, Claude or Codex at zero model cost to the client. Seven chains, one deployment, new capabilities in weeks.

If this sounds like your company: you have the data and still wait weeks for every answer. Put an AI employee on it that shows its work.

Let's get started!











    This site is protected by reCAPTCHA and the Google
    Privacy Policy and
    Terms of Service apply.

    Awesome!

    We’ll be right back for a good talk.
    Cheers ?

    Book a free consultation