Tools we built, working on real data

Alongside client work, we build our own tools and use them on real data. Both are open source, both work in Arabic and English, and both show their working, so you can check every answer.

  • 01

    Humanitarian Reports Tracker

    Reads new situation reports every day, pulls out the figures with AI, and checks each one word for word against the report.

    Gemini or open-source modelBigQueryArabic and EnglishRuns daily

    How it works See the tracker

  • 02

    Local AI Analyst

    An open-source AI model on your own laptop writes the query, runs it, draws the chart and explains the result. Your data never leaves the machine.

    Open-source modelRuns offlineTimesFM forecastsArabic and English

    How it works Try the examples

Humanitarian Reports Tracker

Every figure in a report, with the sentence that proves it.

Reads new situation reports every day, pulls out the figures with AI, and checks each one word for word against the report.

One sentence from an OCHA report, as the tracker stores it

“Since the beginning of 2026, approximately 1,790 Palestinians have been displaced by demolitions for lack of Israeli-issued building permits across the West Bank.”
Indicator
People displaced
Value
1,790
Place
West Bank
Period
1 January to 22 October 2026
Checked against the report
Found word for word

Source OCHA, Humanitarian Situation Report, 22 October 2026

At a glance

  1. 20

    ReliefWeb reports read in its last run

  2. 4

    steps from PDF to a checked table

  3. 0

    figures kept without their source sentence

Gemini or open-source modelBigQueryArabic and EnglishRuns daily

See the tracker See the code on GitHub

How it works

  1. Collect

    New reports arrive each morning from ReliefWeb, your partners' emails, or a shared folder.

  2. Extract

    An AI model reads each report and writes down every figure, its place and date, and the exact sentence it came from.

  3. Check

    Each sentence is searched for in the report. A figure we can't find word for word is marked unverified, so a guess never passes as fact.

  4. Use

    The figures land in BigQuery or a file you own, with a live tracker and a CSV export.

Local AI Analyst

Ask your data a question. Get the answer, the chart and the SQL.

An open-source AI model on your own laptop writes the query, runs it, draws the chart and explains the result. Your data never leaves the machine.

A real answer, recorded from the app on sample data

Forecast households reached per month for the next 3 months

05,00010,00015,000JanFebMarAprMayJunJulAugSepOctNovDecForecastTotal

The forecasted number of households reached in October 2026 is 7,728, with a likely range between 5,972 and 9,664. The forecast for November 2026 is 7,609, with a likely range between 5,723 and 9,698. For December 2026, the forecast is 7,653, with a likely range between 5,701 and 9,838. These forecasts are based on the TimesFM model and are not actual data.

Answered by Qwen2.5-Coder 3B, an open-source model, on a laptop-class CPU; forecast by Google TimesFM 2.5. The aid figures are made up for the demo.

At a glance

  1. 0

    rows of your data sent to the cloud

  2. 6

    real answers recorded from the app

  3. 2

    languages, Arabic and English

Open-source modelRuns offlineTimesFM forecastsArabic and English

Try the examples See the code on GitHub

How it works

  1. Ask

    Load a spreadsheet or CSV and type a question in Arabic or English.

  2. Query

    A local AI model reads the table's columns and writes the SQL. If the query fails, it reads the error and fixes it.

  3. Answer

    The query runs on your computer. You get a table, a chart and a short explanation, with the SQL shown so anyone can check it.

  4. Forecast

    For questions about the future, Google's TimesFM model predicts the next periods and shows the range they will likely fall in.