Skip to content
View omvyas77's full-sized avatar

Block or report omvyas77

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
omvyas77/README.md

Om Hitesh Vyas

I build data and ML systems, and say what the numbers can’t.
M.S. in Information Management, University of Illinois, 2026.

Portfolio     LinkedIn     Email     Hugging Face     Tableau Public


Selected work

finhelm: answers questions about bank SEC filings with citations, or refuses. It refused 18 of the 19 questions with no answer in the filings, finds the evidence 74% of the time, refuses wrongly 12% of the time, and costs $0.03 a question. Python, FAISS, RAG, CI-gated evals, Docker. uplift: a 14M-row randomized test used as the answer key for causal methods. Propensity weighting on the randomized data biased the estimate by minus 91%. Best Qini 0.089, CUPAC cut variance 30.9%, and 3.6% of targeted users were ever shown an ad. dbt, DuckDB, LightGBM, EconML, FastAPI, Dagster.

program-health: seven public sources on Illinois's business school, refreshed and tested on a schedule. Online programs account for 4,781 of the 4,940 students added since 2017. 81% of the rise in A grades happened within courses, 18 schools are compared, and the forecast did no better than a naive one. dbt, DuckDB, GitHub Actions, Streamlit, Tableau. Credit-risk-early-signals: early warnings from consumer complaints filed with the CFPB. LightGBM reaches a PR-AUC of 0.314 on a later period where 0.7% of complaints are flagged; logistic regression 0.286, a fine-tuned transformer 0.121, and 0.124 on a product the model never saw. Python, TF-IDF, LightGBM, transformers, SHAP.


Try them

Project Live Read Code
finhelm Demo on Hugging Face Measuring refusal Repo
uplift Dashboard on Hugging Face Findings Repo
program-health App on Hugging Face
Dashboards on Tableau Public
Brief (PDF) Repo
Credit-risk-early-signals Report (PDF) Repo

A demo that has been idle takes about a minute to wake.


Research


Tools

Data: SQL, Python, PySpark, dbt, DuckDB, Postgres
Modeling: scikit-learn, LightGBM, EconML, PyTorch, Hugging Face
Shipping: FastAPI, Streamlit, Docker, GitHub Actions, Tableau


Contact

Reach me on LinkedIn or at omvyas288@gmail.com.

Pinned Loading

  1. finhelm finhelm Public

    Finance-domain RAG over US bank SEC filings, with a CI-gated evaluation harness. Cited answers, measured abstention, and negative results kept in.

    Python

  2. uplift uplift Public

    Causal experimentation & uplift-modeling platform on 14M randomized incrementality records: RCT-validated observational estimators, ITT/CACE, Qini with bootstrap CIs, and DR off-policy evaluation.

    Python

  3. program-health program-health Public

    An independent look at Gies College of Business program health from public data: enrollment, Big Ten benchmarks, course grades, and Spring 2027 registration. dbt, DuckDB, Python, Tableau Public, St…

    Python

  4. Credit-risk-early-signals Credit-risk-early-signals Public

    Does a consumer's complaint text predict a late company response? CFPB narratives, rolling-window evaluation, calibration and an in-browser scorer

    Jupyter Notebook