I’m Om Hitesh Vyas, a data scientist in New York. I build credit-risk models and retrieval systems, and I measure where they fail.
Five public projects. The code and full results for each one are on GitHub.
Answers questions about US bank SEC filings using only passages it can cite, and declines when the filings don’t say.
18 of 19 unanswerable test questions declined, with 74% of the needed evidence retrieved, at about $0.03 a question.
A randomized experiment on 14 million records, used as an answer key to test which causal methods recover the true effect when randomization is taken away.
20.2% lift in visits once adjusted, against 27.1% from a plain comparison.
An independent look at the Gies College of Business at the University of Illinois, built from public data as a tested dbt and DuckDB pipeline.
4,781 of 4,940 students Gies added since Fall 2017 are in online graduate programs.
A classifier that flags consumer complaints likely to end badly, trained on 679,808 CFPB narratives with a time-based split.
0.29 PR-AUC on the test set, against a 0.71% base rate.
One disparity screen run on two complaint outcomes, to show how much the choice of outcome drives what gets flagged.
65% and 8% of cells flagged on relief rate and on timely response.
Credit risk and forecasting in financial services, and retrieval and customer analytics before that.
Sep 2026 to present
New York, NY
Default-risk modeling: exploratory analysis of borrower, credit-history and loan data that surfaced 20+ predictive risk patterns, SQL feature pipelines that produce 30+ model-ready predictors, and SHAP explanations of what drives default risk.
May 2025 to May 2026
Illinois
Built and validated consumer credit-risk and loss-forecasting models (PD, EAD, expected loss) with out-of-time backtesting. Designed a data-quality framework that monitors 9 critical KPIs across 2B+ records, adopted as the team’s standard pre-deployment check, and built PySpark pipelines on AWS EMR that process 900M+ customer records per snapshot.
Jan 2024 to Jul 2024
India
Rebuilt retrieval for a customer-facing support assistant, raising Precision@3 from 68% to 74% (p = 0.03, paired per-query test). Then ran the confirmatory A/B test (4.2K visitors per arm): self-service resolution rose from 40.0% to 42.4% (p ≈ 0.025), with latency and groundedness guardrails set to roll back automatically.
Aug 2022 to Dec 2023
India
Customer analytics from transaction, basket and channel data: 15+ RFM-based predictors, 6 customer segments with K-Means, segment validation with hypothesis tests, and logistic propensity models to prioritize retention.
A model’s number means little until you know its baseline and where it breaks. That question runs through all of my work.
I finished an M.S. in Information Management, on the data science pathway, at the University of Illinois Urbana-Champaign in May 2026, after a B.Tech. in Computer Science at Nirma University. I work mostly in Python, SQL and PySpark, with scikit-learn, LightGBM, PyTorch, dbt and FAISS.
The fastest way to reach me is email.