Artificial Intelligence and Data Science

Start with who needs data analysis and small AI tools for a real question and why. Around firms, public bodies and start-ups drive demand in Dubai, Abu Dhabi and the other emirates.

• List 10 offers or suppliers in the UAE with price per analysis in AED and who buys.
• Interview 5 people from firms, public bodies and start-ups: what do they buy now, how often, what do they pay, what annoys them?
• Learn three terms: Training and test split (keeping some data hidden to test honestly); Bias (unfair error that falls harder on one group); Data minimisation (collecting only the data you truly need).
• Write a one-page brief: user, need, hoped-for price, and the rule that matters most.

Activity. Complete the market table, five interview notes and the brief.

Standard to meet. Price ten offers, interview five users, and state one need in one sentence.

Know the rule before the work. Main hazards here: bias, privacy breaches and wrong confidence. Hazard thinking means list each step, ask what could go wrong, and decide the control.

• Ask your local authority or free zone which licence, permit or inspection applies. Write the answer and the date you asked.
• For each step (define the question; clean data; build and test; explain; monitor) write the hazard and the control.
• Use a hierarchy: remove the hazard, replace it, separate people from it, change the way of work, then protective equipment last.
• Keep a one-page incident rule: who you call, what you stop, what you record.

Activity. Draw the flow and mark hazard and control at each step.

Standard to meet. Name a hazard and a control for every step, and answer the licence question in writing.

The work is a short chain: define the question, clean data, build and test, explain, monitor. Learn each link, then link them with records.

• Write a checklist for each of the 5 steps: input, action, check, output.
• Time each step three times and note the slowest.
• Record a defect or delay each time it happens: what, where, cause, fix.
• Measure model accuracy on unseen data; this is your main number.

Activity. Run the chain three times (simulated or real) and fill the log.

Standard to meet. Complete three logs, and give the slowest step a cause and a fix.

What you do not measure you cannot improve. Your main measure is model accuracy on unseen data.

• Define model accuracy on unseen data: what counts, what does not, and over what period.
• Draw a simple chart of the last 10 results and mark the average.
• Find the top two causes of loss with a tally sheet, then fix the biggest one first.
• Check measuring tools: are they accurate, clean and in date?

Activity. Collect a tally for one week and make a Pareto chart.

Standard to meet. Complete the tally, name the biggest cause, and plan one fix with a date.

Know the cost of one analysis before you set the price. Include materials, labour, energy, transport, licence share, waste and a safety margin.

• Cost per analysis = total cost divided by analyses sold (not made).
• Margin = (price minus cost) divided by price.
• Break-even units = fixed costs divided by gain per unit.
• Test the price with two users and write down the answer.

Activity. Build the cost sheet with AED on every line and have your mentor check it.

Standard to meet. Check that no cost is missing, and calculate the margin twice by different routes.

Customers return when you keep promises. Promise only what your workflow can deliver.

• Write a one-page offer: what is included, what is not, price, time, how you fix mistakes.
• Write a short script for the first call or visit.
• Agree dates in writing; confirm each change.
• Handle a complaint in four steps: listen, apologise for the impact, fix, record what changed.

Activity. Present the offer to two users and note changes.

Standard to meet. Write an offer with scope, price, time and a fix rule, and collect comments from two users.

Small, real and recorded. Change one thing at a time and learn quickly.

• Keep a daily log: date, units, cost, delays, defects, safety notes.
• Hold a 15-minute review each week with your mentor.
• Write down every complaint and what you changed.
• Take photos only where allowed and never of people without consent.

Activity. Complete 30 days of the log and four weekly reviews.

Standard to meet. Log 30 days, give each defect a cause, and keep weekly notes.

A decision beats a feeling. Compare plan with result and choose to scale, change or stop.

• Make a before and after table: cost per unit, defects, delays, repeat users.
• Write three lessons.
• Decide scale, change or stop, with the numbers behind it.
• Give two peers one strength and one fix each.

Activity. Present to a review room and collect one written comment from a user.

Standard to meet. Justify the decision with numbers, and give two peer reviews.

A model that scores well on the data it learned from may fail on new data. An honest test keeps some records hidden until the end.

• Split the records before you start: for example 70% to train, 15% to tune and 15% to test. Never look at the test part while building.
• Pick a measure that fits the question: accuracy for balanced groups, but precision (how many flagged cases were truly positive) and recall (how many true positives were found) when one outcome is rare.
• Compare the model with a simple baseline, such as always predicting the most common answer. A model that cannot beat the baseline is not useful.
• Write down the split, the measure and the result before you change anything, so you cannot fool yourself later.

Activity. Take a sample table of at least 500 rows from your mentor, split it, run a simple model and a baseline, and report both scores.

Standard to meet. Write down the split sizes, show the baseline score next to the model score, and leave the test part untouched until the final run.

Data about people can harm them if it is unfair, leaked or used for a purpose they never agreed to. Checking this is part of the job, not an extra.

• Check error rates for each group, such as age band or nationality group. If one group is wrong far more often, the model is unfair even when the overall score is high.
• Collect only the fields you need. Remove names and ID numbers before analysis where you can, and store the key to re-link them somewhere separate.
• Know the purpose: data collected for one reason should not be reused for another without permission and a lawful basis.
• Keep a short record of where the data came from, who may see it and when it will be deleted.

Activity. Calculate error rate per group from a sample results table and write a half-page note: fair enough to use, or not, and why.

Standard to meet. Show the error rate for every group, justify the data fields one by one, and give a clear recommendation with reasons.

A good result is only useful if the decision-maker understands it and if it keeps working after launch. Data changes, and models slowly go stale.

• Lead with the answer in one sentence, then give the number, then the limits. Avoid jargon; say 'wrong about 1 time in 10' instead of a technical score.
• Show uncertainty: give a range, not a single figure, and say how many records it rests on.
• Monitor after launch: track the live error rate each week against the test result and set a trigger, for example retrain if it drops more than 5 percentage points.
• Keep a log of model versions, data dates and who approved each change.

Activity. Write a one-page summary for a non-technical manager using the results from module 1, with one chart and one monitoring rule.

Standard to meet. Answer in the first sentence; range or limit stated; a numeric monitoring trigger set.

On the slide: L1 Basic · What data analysis is for

Notes: Name the five steps of a data job (question, clean, build and test, explain, monitor) and give one real question a firm could ask. Notes: use a small table of ten rows; learners point to one row and one column.

On the slide: L2 Awareness · Bias, privacy and lawful data

Notes: Explain what bias is, why data is minimised and why only lawful data is used. Notes: show one field that is not needed and remove it; teach that data collected for one purpose is not reused for another without permission.

On the slide: L3 Foundation · Split data and test honestly

Notes: Apply a split of 70% train, 15% tune and 15% test and compare the model with a baseline. Notes: check the practice case of 2,000 records split 1,400, 300 and 300; 288 of 300 right is 96.0% but the baseline of 95.0% leaves a gain of only 1.0 point.

On the slide: L4 Practitioner · Run a full analysis and explain it

Notes: Run the whole workflow on a sample table of at least 500 rows and write a summary for a manager. Notes: answer in the first sentence, give a range, state how many records it rests on and set a numeric monitoring trigger.

On the slide: L5 Specialist · Test fairness across groups

Notes: Analyse error rates for each group and decide whether a model is fair enough to use. Notes: 30 errors in 500 for group A is 6% and 75 in 500 for group B is 15%, a gap of 9 points, so review before use; price check is AED 9,000 per analysis against AED 3,600 cost, a margin of 60.0%.

On the slide: L6 Advanced · Coach a team on model checks

Notes: Set a baseline rule, a fairness check and a weekly drift review for a small analyst team. Notes: retrain if live accuracy drops more than 5 points; sample five analyses a week for split sizes written down and the test part untouched.

On the slide: L7 Super expert · Design the whole data and AI system

Notes: Design the data rules, testing, fairness, monitoring and approval log for several analyses, audit privacy risks and write a 12-month plan. Notes: plan for 12 analyses a month; present one risk, one saving and quarterly targets for accuracy on unseen data.