The short version

You design and execute marketing analytics models, and you like doing it.

Simply: you are a modeller. Every day that means MaxDiff and TURF, key driver analysis, cluster and latent class segmentation, conjoint and other choice models. You write the code, you run the estimation, you check the output, and you work with our client services team to shape the story the client needs. You will lean on AI heavily to do all of it faster and better. If this reads as a description of your actual job rather than of a job you have supervised, keep reading.

What you will personally build

At any point we have eight to ten studies running with an analytics component. Three or four are in heavy analysis at the same time, and those are yours. In a normal week you will:

- Specify and estimate a MaxDiff, run TURF on the output, and turn it around to the client services team on your own authority
- Build a key driver model, separate what is actually driving the outcome from what is collinear noise, and say what it means for the decision the client is making
- Prepare choice data and estimate a conjoint, working with the analytics lead on design and specification
- Run a segmentation, defend the number of segments you landed on, and profile them on variables you deliberately held out of the clustering
- Build the typing tool that goes with it, so the client can assign new respondents to the segmentation after the study closes
- Write the code yourself, in R or Python, and be able to explain any line of it
- Use AI to get through all of that faster, with the verification discipline that makes the output trustworthy, and build the reusable pipelines so the next one is quicker still
- Keep your own view of which studies are coming, when data lands, and what each one will need from you

What this role is not Being explicit here saves everyone time, ours included.

- Not a data processing role. You will prepare and validate analysis-ready datasets for your own models, which is real work, but production data processing is not part of this job. (Weighting, tabbing, significance testing on banner tables, crosstab production and survey QA belong to a separate department at client)
- Not a management role. No direct reports.
- Not the senior analytics seat. Thought leadership, methodology partnership with our Chief Research Officer, leading client calls and owning major readouts sit with our analytics lead. There is a path forward, but that is not this role on day one.

What we need

Must have:





- 5 to 8 years of hands-on market research analytics or marketing science. Doing it yourself. Not commissioning it, not managing the people who do it, not interpreting output somebody else produced.
- Personal execution of MaxDiff, TURF, key driver analysis, cluster and latent class segmentation, and conjoint or other discrete choice models. You should be able to walk us through a study you built end to end and say clearly which parts were your hands on the keyboard. We are weighting choice modelling and latent class segmentation especially heavily, because that is where we most need depth in this seat.
- Working fluency in R and Python for statistical modelling. You choose your stack, and we care that the work is right and that somebody else can rerun it.
- Clean, repeatable, documented code. You will build the same kinds of models many times. We would like the fifth one to be faster than the first.
- The judgment to structure a model and interpret its output so client services can build a clear, concise C-suite narrative on it. You are not writing that narrative. You are making sure it can be written, and that it holds up.
- The ability to align a modelling approach with the client's business objective, partnering directly with our US-based Research VPs. When the strongest driver turns out to be something the client cannot change, we want you to notice and say so rather than reporting it and stopping.
- Enough concurrency to hold three or four live analyses at once. Negotiating timing directly with several project owners, and raising a conflict early rather than working the weekend in silence.
- Machine learning applied to survey-scale data. Regularized regression (ridge, lasso, elastic net) for driver models where the predictors are collinear, which survey batteries almost always are. Tree-based ensembles for variable importance and for building segment classifiers. Cross-validation and holdout discipline as a habit rather than an afterthought. Just as important: knowing when not to reach for ML,



because a 400-respondent driver model a client has to act on usually needs an interpretable coefficient more than it needs a better fit statistic.
- A quantitative degree: statistics, mathematics, economics, marketing science, or a quantitative social science. A master's is strongly preferred. If you do not have one, show us the equivalent: Sawtooth certification, published work, or a portfolio of choice models you personally specified and estimated.
- Comfortable working US Eastern hours from India, communicating day to day over Slack.

Nice to have:

- Hierarchical Bayes written from scratch. Estimating a conjoint through Sawtooth or an equivalent package is a must-have above. Writing the sampler yourself, in Stan, bayesm or equivalent, is the higher bar and a real differentiator.
- Predictive or regression-based work in a market research context.
- Experience explaining or presenting statistical findings directly to clients.
- SPSS. Helpful for reading legacy work. Not something you would use here daily.

On AI, which we mean seriously We flagged this at the top and it is worth being precise about, because most job descriptions ask only for openness to AI tools and we are asking for something more specific. It is one of the real differentiators in this work right now. Two things we want, and one bonus.

- You use it to write, debug and check code faster, with verification discipline. Holdout tests, reconciliation against results you already know, sanity checks on parameters that should move a particular direction. AI-assisted is fine and we do it too. AI-assisted and unverified is not, and we will ask you how you check.
- You build reusable, parameterized pipelines. So that the fifth MaxDiff takes a fraction of the time the first one did, and so that the analysis is consistent across studies rather than rebuilt from memory each time.
- Bonus: you experiment at the edges. You have pushed on what LLMs can and cannot do inside research analytics, and you have a considered view of where they hold up under verification and where they fall over. We would rather hire someone who has tried things and can tell us exactly what broke than someone who has only read about it.

How we will assess this A short work sample. We will give you a dataset and a real question, you will have a few days, and then we will talk through what you built and the choices you made along the way. It is not a trick, and it is not unpaid production work. It is the fastest honest way for both of us to find out whether this fits, and it matters more to us than anything on your CV.

📌 Marketing Science Analyst (India)
🏢 Talentra Global Group
📍 India