Measure first. Then build.

A Durban AI and software studio. We find where your business loses time and money, then build the fix and test it on your own data.

Book an audit See if AI would pay off
Fixed-price audit
R35,000
Written answer in
2 weeks
Done by
The founder

Click the water to ping it. Seabed from NOAA's ETOPO 2022 model, height exaggerated. The current is illustrative.

Agulhas means needles.

The usual explanation: off this cape, early Portuguese compass needles pointed almost true north.

We expect the same accuracy from software.

Is it worth automating?

Put in your numbers to compare doing the work by hand with AI plus a person checking it. Sometimes the answer is no.

These are made-up examples. Replace them with yours.

More options

Use your own export

Choose a CSV of invoices, orders or tickets to count the monthly volume and check for duplicates and blanks. The file is read in this tab and never uploaded.

Worth an audit.

Even the cautious estimate pays back within 18 months. The audit checks whether the accuracy holds on your documents.

By hand
R72,000 a month
With AI
R44,000 to R55,000 a month
Pays back in
4 to 13 months
Send us these numbers

A modelled estimate from made-up examples, not a measurement of your work.

See the working
Goes through unchecked
52%
Wrong entries a month
30 to 120
Audit plus build
R90,000 to R210,000
Monthly cost at each confidence setting. Too low and errors cost you; too high and people still do the work. The dashed line is doing it all by hand. The shaded band is half to double the error rate.
240 synthetic documents, by model confidence. Right of the line goes through unchecked. The amber ones are wrong.

What this assumes.

  • The accuracy curve comes from 5,000 synthetic documents generated in your browser, not from a client or a live model.
  • The range runs from half to double its error rate.
  • The model's confidence is assumed to be honest, and checking a document is assumed to take as long as doing it. Real models can be over- or under-confident, and your real curve is what the audit measures.
  • Build prices are the published document automation range.
  • Your people are assumed to make no errors, which flatters the hand process.
  • The verdict is our rule of thumb: both ends of the range inside 18 months is worth an audit, more than 24 months on the best case is not.

What we do.

Clear prices, and every build tested on your data before anyone relies on it.

Start with the AI Operations Audit

We study your real process, then tell you in writing where AI will save you money and where it won't.

R35,000 Fixed price, 2 weeks. US$4,900 abroad.

See what you get

All services, scope and pricing

Proof you can check.

Six public builds with tests you can run. Each one says where it falls short.

  1. Support answers with sources

    Answers from a help centre with cited snippets, or opens a ticket when unsure.

    Tested on
    23 invented documents, 144 tests
    Result
    Right source in the top five for 22 of 22 hand-written questions.
    Where it falls short
    It checks a source is cited, not that the source backs the answer. 3 of 8 questions outside its documents still got an answer.
    RAG Support Bot on GitHub

  2. Spot customers about to leave

    Ranks customers by how likely they are to leave, and forecasts monthly recurring revenue.

    Tested on
    IBM's public Telco sample of 7,043 customers and a synthetic series, 43 tests
    Result
    75.9% of the riskiest tenth did leave.
    Where it falls short
    It tests the method, not your customers. The revenue forecast missed by 12.65% on its synthetic test.
    Churn & MRR Forecasting on GitHub

  3. Recommend for new and regular users

    Mixes item details with what similar users chose, depending on how much history a user has.

    Tested on
    A synthetic catalogue and MovieLens-100K for 109 users, 54 tests with 1 skipped
    Result
    Reaches 97% of the catalogue on the synthetic set, against 13% for a most-popular list.
    Where it falls short
    A simple most-popular list still beats it on accuracy, on both datasets.
    Hybrid Recommender on GitHub

  4. Reuse earlier AI answers

    Reuses a stored AI answer when a new question is similar enough to an earlier one.

    Tested on
    30 synthetic support questions, 176 tests with 1 skipped
    Result
    At the 0.80 setting, 1 in 30 reworded questions reused an answer and 1 in 30 near-misses wrongly got one.
    Where it falls short
    That setting alone isn't safe: at 0.50, 70% of near-misses wrongly got a reused answer.
    Semantic Cache on GitHub

  5. Ask a database in plain English

    Turns a plain question into a database query, and checks the query only reads data before it runs.

    Tested on
    A synthetic data warehouse, 62 tests
    Result
    No accuracy figure yet.
    Where it falls short
    We haven't measured how often its answers are right.
    InsightEngine on GitHub

  6. Compare ways of reasoning

    Tries five ways of getting AI to work through a problem and scores each one.

    Tested on
    40 public maths problems, answered by a simulated model, 180 tests
    Result
    No real-model result yet.
    Where it falls short
    Not yet run on a real AI model, so it says nothing about real models.
    Thinking Loop on GitHub

Figures are from each project's public notes, read on 2 November 2026. Full archive on GitHub

You work with the founder.

Darrshan Govender, founder and senior AI engineer. The person in the pitch is the person writing the code, from first call to launch.

  • No handoffs

    No account managers, offshoring or junior developers.

  • Based in Durban

    On site in KwaZulu-Natal, remote elsewhere.

How it works.

  1. Measure

    We learn your business and agree the numbers the work will be held to.

  2. Design

    You review every screen before we build it.

  3. Build

    A live demo every two weeks.

  4. Launch and support

    We handle go-live, monitoring and fixes, with a clean handover whenever you want it.

Common questions.

Can AI help my business?

Probably, but only where it earns its place. Real wins are usually manual back-office work and getting answers out of your own data. We'll tell you if AI isn't the right tool.

Where should we start?

Usually with the AI Operations Audit: two weeks, a fixed R35,000 (US$4,900 outside South Africa), and a written answer on where AI will save or earn you money and where it won't. If you already know what you need, we can go straight to a build.

How do you decide if automation is worth it?

We compare the cost by hand with the cost of AI plus a person checking it, including wrong entries and running costs. The calculator on this page does that sum. The audit measures the real accuracy on your documents.

How long does a project take?

The audit takes two weeks. Most builds take three to ten weeks. You get a timeline before any work begins.

What does it cost?

The audit is a fixed R35,000. Websites start at R65,000, and most AI builds fall between R55,000 and R250,000. Outside South Africa the audit is US$4,900 and builds start at US$7,500. Larger platforms are phased, and you get a fixed quote before work begins.

Who writes the code?

The founder, with no offshoring and no junior handoffs. If a larger project needs a specialist, we agree it with you in writing first.

Do you work outside South Africa?

Yes, across Africa, the UK, Europe, the Gulf and North America. We're on UTC+2 all year, with overlap with Europe through the day and with US East Coast mornings.

Let's find the leak.

Tell us what you're trying to fix. You'll hear back within one working day from the person who would do the work.