Data Scientist, Response EvaluationJob DescriptionAbout UplevylUplevyl builds AI-powered knowledge and community infrastructure for organizations serving women. Our products include UpGenie (our domain-specific AI assistant), WeHub (our community platform) and UpSocial (a social platform for women). We work with mission-driven partners to turn complex, high-stakes information into transparent, trustworthy guidance, and to build systems that create lasting impact.

We hold a simple conviction: in high-stakes domains like rights, law and financial security, a generic AI is not enough. The answers people stake their lives and livelihoods on need a purpose-built system with verified, native data. That is what we build, and it is why the work is urgent. AI is reshaping how the world learns, works and earns, and the women we serve cannot afford to be left behind.

We have also made a deliberate choice about how we build: a small team of exceptional people, paid well above market, each doing work that would normally take several. We would rather be ten people who move the world than thirty who move paper. That choice sets the bar for every hire, including this one.

About the RoleThe Data Scientist, Response Evaluation, owns the math underneath UpGenie. UpGenie answers high-stakes questions about rights, law and safety, and you will decide how it scores and ranks evidence, how we measure whether an answer is right, how confident the system needs to be before it answers, and how we balance accuracy against latency and cost.

The answer to a rights question can change at every level of government: federal law, the 50 states and DC, and about 90,000 local governments, which the US Census Bureau counts as counties, cities, towns, and school and special districts. Our system retrieves and ranks evidence across federal, state and local law in several domains, and our evaluation results drive real product decisions. You will make sure those decisions rest on things we can reason about: scoring functions whose behavior we understand, metrics with confidence intervals, thresholds with stated guarantees, and experiments with enough power to separate signal from noise.

The role suits someone trained in mathematics, statistics or a quantitative science who has turned that training into working software. You can derive an estimator and prove its properties, then implement it, test it on real data and ship it.

You will ship to production from your first weeks, with real users and real partners depending on what you build.

What You’ll Own1. Ranking and Retrieval Math The scoring functions behind search and ranking: lexical scoring, embedding similarity, hybrid score fusion, learning-to-rank, re-ranking and diversity.



You will know the geometry of the embedding spaces we use and where it breaks down.

2. Evaluation Statistics

Metric design, sampling plans, confidence intervals, significance and power, inter-annotator agreement, and correction for noisy automated judges. Every number we report comes with its uncertainty.

3. Calibration and Abstention

Confidence scores that mean what they say. You will calibrate model scores, set abstention and escalation rules with explicit error guarantees, and decide when the system should ask a follow-up question instead of answering.

4. Optimization and Model Improvement

Objectives and loss functions for fine-tuning and ranking models, principled hyperparameter and prompt search, and a clear model of the trade-off between accuracy, latency and cost, so product decisions rest on numbers.

5. Monitoring and Data Quality

Catch quality changes as they happen: shifts in query and document distributions, embedding drift across model versions, change-points in production metrics, and duplicate or near-duplicate content at scale.

6. Decision Support

Short, precise analyses that product, engineering and partners can act on. Lead with the answer, show the math, then say it in plain English, including what the data does not support. No false precision: 92.37% on 150 cases is not a finding.

Who We’re Looking ForWe hire across levels. Depth of understanding and results you can show matter more than titles or years of experience. We will ask you to solve problems and build during our process, because we have learned that knowing the math and applying it to messy, real systems are different skills.

Deep Mathematical Foundations

Probability, statistics, linear algebra and optimization at the level of derivations and proofs, not just library calls. You can tell when an assumption is doing the real work in an argument.

You Ship Working Code

You are fluent in Python with NumPy, SciPy and pandas, and you work in PyTorch or JAX. Your analyses are reproducible and tested, and they run in production, not only in a notebook.

Statistical Judgment

You know when a difference is noise. You design the experiment before you run it, and you report uncertainty by default.

Rigor In our domain, a confident wrong answer can harm someone. You test before you claim, you measure before you optimize, and you flag what you are unsure about.



You keep facts, inferences and assumptions separate.

High Agency

You unblock yourself. Given an ambiguous problem with no owner in sight, you scope it, propose an approach and build.

A Clear Communicator

You can explain a confidence interval to a non-technical partner and defend a modeling choice to a senior researcher, in plain words either way.

Even Better If

- You have a graduate degree or research experience in mathematics, statistics, theoretical computer science, physics or a related field.
- You have worked on information retrieval, ranking, recommender systems or search quality.
- You have built evaluation, experimentation or A/B testing infrastructure for machine learning systems.
- You have used Bayesian methods, causal inference, conformal prediction or sequential testing on real problems.
- You have worked with Vertex AI, GCP or comparable cloud ML stacks.
- You have public work we can look at: papers, open source, technical writing or competition results.

Before You ApplyWe want to be direct about what this is. This is a high-bar role on a small team, with real ownership and high expectations. Because our products work in high-stakes domains, we value rigor over shortcuts. Speed matters, but so does knowing when an elegant model or a plausible-looking result needs a closer look. If you want a pure research role, or a defined lane with clearly separated responsibilities, this is not the right fit, and that is fine.

If you want to bring mathematical rigor to a system people depend on, we want to hear from you.

Our interview process is thorough, and we will walk you through every step of it in our first conversation.

How We Work at UplevylWe finish what we start, and we do it well.

We follow through on what we say we’ll do, and we care about the difference our work makes.

We listen before we decide.

Before we act, we ask who it actually affects: a customer, a partner, a teammate or the communities we serve. Trust here is built the plain way, by consistently showing up for people.

We move with ownership and urgency.

We don’t wait for perfect information or for someone else to raise their hand. If something looks like it could go wrong, we say so early. We make the call, we move fast, and we hold a high bar for quality.

We are better together than alone.

We work across teams, say what we actually think, and go out of our way to help each other succeed. Leadership here is measured by the impact you create.

We stay curious and keep learning, including how to use AI well.

We hold ourselves to the bar we’re building toward: questioning our own assumptions and using AI to think better and move faster.

📌 Data Scientist, Response Evaluation (Noida)
🏢 Uplevyl
📍 Noida