Writing
Notes on applied AI and the distance between a working demo and a system a business can depend on.
01Fourteen months of the same weekFourteen months at A2SV — five sessions a week, a contest on the last day, and the same hardest moment in nearly every week of it. What I came out with was not a set of algorithms but a question: not whether the code runs, but what it costs to run, which is also how you judge code you did not write.02A high number is not evidenceFive months teaching the data-mining labs at Nile University. A very high accuracy is usually a fact about the experiment rather than the model — leakage, cleaning done before the split, repeated rows — and the reaction it deserves is suspicion, not celebration.03The broken process behind the requestA chatbot is the part of the work a client can picture, so it is the part they ask for. One retailer asked for one; reading the messages their customers had actually sent turned up scattered product information, an unclear returns policy and no way to follow an order — none of it a conversation problem.04The deadline is part of the answerWe spent three months building a conference recommender that worked. Its suggestions were relevant and most of them were useless, because recommending a venue is a freshness problem before it is a matching problem.05Cleaning the noise, not the languageWhat I learned building a 330,000-row Arabic sentiment dataset from scratch — dialects, sarcasm, and why the decisions you make while cleaning are part of the data.