Umut Altun

I’m a senior data scientist in Rome. For 7+ years I’ve built data systems that run in production — LTV and subscription prediction, marketing analytics, and lately agentic AI / LLM systems — for products reaching well over a billion users. Currently at TapNation (mobile gaming) and Wavera (subscription social); before that, telecom AI at P.I. Works. MSc Data Science, Sapienza Rome.

This is where I write about the work: LTV modeling, experimentation, and building LLM systems that survive contact with production. More on about, or see my CV.

Writing

Designing an LLM system that thinks in Turkish

Building a Text2SQL agent in Turkish turned out to be five percent about the model and ninety-five percent about everything around it. A note on staying in-language end to end, the two-byte bug that shredded every export, and the English-by-default tax.

May 2026

The eight silent seconds that made my app feel broken

The answer was correct and arrived in eight seconds. Users still thought it was broken — because for those eight seconds, nothing moved. A note on perceived latency, server-sent events, and the proxy setting that quietly undoes all of it.

April 2026

Letting the LLM fix its own SQL — but only twice

Feeding a failed query’s error back to the model so it can correct itself feels like magic the first time. The third time it ‘fixes’ the same query, you realize you’ve built a way to set tokens on fire. On bounded self-correction and the dry-run that runs before any query does.

March 2026

Teaching an analytics agent when not to ask

Getting an agent to ask for clarification instead of guessing is the easy half. The hard half is teaching it to stop asking — because an agent that interrogates you on every turn is one nobody uses.

January 2026

Two tenants don’t need multi-tenancy

When a second customer arrived, I reached for the textbook shared-schema multi-tenant architecture — and then noticed it would make data isolation depend on an LLM getting a WHERE clause right every single time. On choosing the architecture the problem needs, not the one that looks good on a whiteboard.

November 2025

The obvious fix for a hallucinating SQL agent is the wrong one

An LLM that invents a filter value doesn’t throw an error — it quietly returns the wrong answer to someone who can’t read SQL. The intuitive correction makes it worse. On value grounding, conservative canonicalization, and tuning for the expensive mistake.

October 2025

The router that’s allowed to say ‘I don’t know’

My first Text2SQL agent was one big prompt holding the whole schema. It confidently joined tables that had no business being joined. The fix wasn’t a better prompt — it was decomposition, and giving the routing step the right to refuse.

September 2025

MMM wants two years of data; I had two months

Marketing mix modeling is famously data-hungry, and most real situations can’t feed it. The professional move isn’t to force the model to produce a confident number anyway — it’s to build it to know when it has nothing useful to say, and admit it.

June 2025

One model per country, and the tax I paid for it

A single global marketing mix model averages the US and Brazil into a country that doesn’t exist. Splitting it per country fixes that and immediately starves each model of data. On choosing which problem you’d rather have.

April 2025

iOS broke attribution, so I stopped attributing

After Apple’s privacy changes, per-user attribution on iOS quietly became fiction. The instinct is to patch the broken signal. The better move is to switch to a method that never needed user-level tracking in the first place — and measure spend, not installs.

January 2025

The installs you can’t attribute

Attribution tools can only credit the install they directly touched. But paid spend also drives installs nobody clicked an ad for — word of mouth, chart rank, the friend who saw you playing. Measuring that halo, not paid ROI, is what I pointed a marketing mix model at.

November 2024

Marketing isn’t linear, and neither is my model

Two assumptions quietly wreck a marketing mix model: that a dollar spent today only matters today, and that the millionth dollar works as hard as the first. Adstock and saturation are how you fix both — and why MMM is a curve-fitting problem, not a regression.

September 2024

Why my marketing model has opinions before it sees data

Regression-based marketing mix modeling has a dirty secret: the data doesn’t actually pin down the answer. Several wildly different attributions fit equally well. The fix isn’t more data — it’s admitting you have prior beliefs and writing them down.

July 2024

A thousand simulations per cohort, and never a loop in sight

Monte Carlo confidence intervals have a reputation for being too slow for production. They’re not — naive Monte Carlo is slow. The fix is to stop thinking one simulation at a time and let the array library do all thousand at once.

June 2024

Two LTV models that disagree, and the rule for which to believe

One model is precise when a cohort has lots of data and falls apart when it doesn’t. The other is robust on sparse data and crude when there’s plenty. The valuable code isn’t either model — it’s the dozen lines that pick between them.

March 2024

Old cohorts are lying to you

The more historical data I gave the model, the worse it predicted. That isn’t supposed to happen — unless the thing you’re modeling quietly changes underneath you, which in mobile gaming it does, constantly.

December 2023

Why my LTV confidence intervals sample from two different distributions

A point estimate is the wrong output for a budget decision. Building honest confidence intervals meant sampling retention from a Beta and revenue from a Gamma — because the distribution you draw from is a statement about what kind of quantity you’re modeling.

September 2023