Hello, I’m Nripesh.

The path here

I studied mathematics and statistics at Grinnell College, with a detour through the Budapest Semesters in Mathematics. I started out in data-science consulting, building predictive analytics for Fortune 500 retailers, where I learned to turn a vague business problem into something a model can actually answer.

In 2021 I joined Chipper Cash, a cross-border payments company, as a data scientist. The job was supposed to be analytics. It quickly became infrastructure. I led the company’s move to dbt and designed the core financial data models that still power its analytics, dashboards, and internal tools, and I ended up running the data-transformations practice that more than fifteen engineers build on.

From there the work moved steadily closer to the things that cost real money. As a staff-level tech lead on risk I designed and shipped the fraud-detection stack: a model that scores every new user, real-time transaction scoring, account-takeover defenses. Today, as Head of Risk Intelligence and Automations, I lead a small team building ML and LLM systems across seven markets, replacing vendor black boxes with systems we can open up, measure, and trust.

What I’ve built

Most of my favorite work has the same shape: take something slow, manual, and vendor-dependent, and turn it into a system that is fast, owned, and measured. An LLM pipeline that adjudicates watchlist alerts, evaluated against a double-blind golden dataset before it touched a single real case, then run over a 148,000-alert backlog at four cents a decision. A fraud model that blocks six thousand bad signups a day. A pipeline that reads partner fraud reports with an LLM and acts in twenty seconds instead of thirty minutes. A platform that replaced a six-figure vendor contract. The pattern I care about is the middle step: you do not ship a model because it demos well, you ship it because you built the eval that proves it.

How I think about building

• An LLM system without an eval is a demo. Build the golden dataset before you build the pipeline.
• Get the data layer right first. Everything downstream inherits its quality.
• Prefer systems you can open up. A vendor black box you cannot inspect is a liability dressed up as a convenience.
• Automate the judgment, not just the clicks. The value is in encoding the decision, not scripting the busywork.
• A noisy alert is worse than no alert. Every false positive spends trust you do not get back.
• Write it down. Explaining a system clearly is also how you find its flaws.

Outside the work

I think a lot about risk and uncertainty, which spills over into chess and poker. I follow football closely and recently fell in love with pickleball. And I read seriously and eclectically, and write essays on the books I keep turning over, because explaining a book clearly turns out to be a lot like explaining a system. It is how you find out what you actually think.

Right now I am deep in LLM evals and the question of when to trust a model with a real decision. If you are working on the same problems, I would like to hear from you.