Binju Karki · Data Analyst & Engineer

Ground truth

Every role I've had has had the same catch: get the data wrong, and it's not an abstraction that breaks: it's a real person on the other end of it.

At Cedar Gate, that meant HIPAA-bound healthcare records. At GVSU's Charter Schools Office, it's FERPA-bound student records for 78 schools: data that gets validated and suppressed before it ever reaches a dashboard, because the person behind each row is a kid. Before that, as a Study Abroad Peer Advisor, the "data" was a student in front of me deciding whether to spend a semester on the other side of the world.

That's the thread running under everything I build now, including the AI agents: I don't let a model, a report, or a dashboard say something it can't back up.

2023

Healthcare Data, Cedar Gate · HIPAA & PHI

2025

Study Abroad advising · Real students, real stakes

2025–26

Student Data, GVSU · FERPA & PII, 78 schools

2026

AI Agents · Grounded, not guessing

Projects & results

What I've built, what it does, and what came out of it: from a BI platform in progress for 78 schools to deployed AI agents and statistical research.

R · Power BI · DAX · Power Query · Star Schema

K-12 Data & Reporting Platform

Building now

I own the data pipeline behind GVSU's Charter Schools Office reporting: I've already cleaned, integrated, and validated K-12 data across all 78 schools, and I'm now building the interactive Power BI dashboard that unifies compliance, financial, academic achievement, academic growth, enrollment, and peer-benchmarking reporting into one place per school.

Completed: data cleaning, integration, and validation across enrollment, assessment, and financial sources for all 78 schools. In progress: a unified interactive dashboard covering all six reporting domains for all 78 schools at once.

schools, cleaned data & dashboard in one place

reporting domains unified: compliance, financial, achievement, growth, enrollment, peer benchmarking

R + Power BI

cleaning pipeline & dashboard build

Platform architecture

Diagram reflects the pipeline architecture I built and am building; individual student-level data is FERPA-protected and never shown.

Agentic AI · Streamlit · OpenAI

Research Synthesis Agent

Live

An agentic research assistant that reads uploaded papers or abstracts and produces a structured literature-review synthesis (themes, methodologies, conflicting findings, and research gaps) instead of a generic summary.

Result: a working, publicly deployed agent that autonomously decides when to pull outside research via tool-calling, rather than only summarizing what's given to it.

custom retrieval tool

structured output fields

GPT-4o-mini

model

Live app ↗ Source code ↗

Agentic AI · FastAPI · Streamlit · OpenAI

MentorMate · AI Study Companion

Live

An agentic study assistant for graduate coursework that decides, per question, whether it needs a term definition, course-note context, or a study plan before it answers.

Result: 100% correct autonomous tool selection across 8 evaluation cases: the agent reliably picks the right tool instead of guessing or hallucinating an answer.

orchestrated tools

100%

tool-selection accuracy, 8 eval cases

deployment paths (FastAPI + Streamlit)

Live app ↗ Source code ↗

R · tidymodels · Multinomial / LDA / QDA / Poisson

Family Involvement & Student Grades

A classification study on the NCES Parent & Family Involvement in Education survey, predicting a student's usual letter grade from parental engagement and household background.

Result: multinomial logistic regression was the best of 3 models tested, but with a majority-class baseline of 56.5% (most students get an "A"), its 58.9% CV accuracy is a modest lift, not a strong classifier. The more useful finding: family involvement (parent-teacher conferences, activity attendance) was a statistically significant predictor of grade category even after controlling for household income.

13,417

households analyzed

models compared, 10-fold CV

58.9%

best CV accuracy

Distribution of parent-reported usual grades

School activity attendance by usual grades

Confusion matrix for final multinomial model

Residuals vs fitted values for Poisson regression

R · Monte Carlo Inference · ggplot2 · Leaflet

U.S. State-Level Census & Election Analysis

Merged state-level census and presidential election data (2008–2016) to test whether income, education, and housing cost differ between states won by each party.

My contribution: ran a 5,000-iteration permutation test showing education levels differ significantly between Democrat- and Republican-won states, and a 10,000-sample bootstrap producing a 95% CI of $55.6K–$57.6K on national median income.

5,000

permutation iterations

10,000

bootstrap samples

election cycles compared

Median income by party when the party won the state

Total presidential candidate votes by decade

Density of median monthly rent when the party wins the state

State win count heatmap by income bracket

R · tidymodels · 8 Modeling Techniques

STA 631 Statistical Modeling Portfolio

A self-directed modeling portfolio built for STA 631 (Statistical Modeling), covering 8 techniques end-to-end in R using the tidymodels framework: multiple linear regression, polynomial regression, interaction models, multinomial logistic regression, LDA, QDA, Poisson regression, and ridge/lasso penalized regression, applied to the Auto and Diamonds datasets with full diagnostics and cross-validation.

Result: compared 3 classification approaches for predicting a car's origin (up to 78.8% test accuracy), and 2 penalized regression approaches for predicting mpg, with lasso outperforming ridge (RMSE 3.38 vs. 3.52). An interaction model on the Diamonds dataset reached R² = 0.998.

modeling techniques, one portfolio

0.998

R² on best-fit interaction model

78.8%

best classification test accuracy (QDA)

Correlation matrix and multicollinearity check, Auto dataset

Quadratic polynomial fit of mpg vs horsepower

Log-carat vs log-volume by cut, Diamonds interaction model

Cylinders vs key predictors, Poisson regression EDA

Full portfolio on Rpubs ↗

D3.js · Observable · Interactive Visualization

A Visual Analytics Approach to Wine Quality

Ten interactive D3.js visualizations exploring what physicochemical properties drive wine quality, built for wine producers, sommeliers, and enthusiasts to explore directly.

Result: identified alcohol content and volatile acidity as the two strongest quality predictors, and shipped a working blend-simulator so users can test the finding themselves.

6,497

wines in dataset

interactive visualizations

physicochemical features

Distribution of red vs white wine donut chart

Correlation heatmap of wine properties and quality

Alcohol content vs wine quality scatter plot

Interactive slider predicting wine quality

Parallel coordinates plot comparing red and white wine

Radar chart simulating red-white wine blends

Skills

Languages & Data Engineering

PythonRSQLDAX Power QueryPySparkDatabricks Amazon RedshiftMySQLAmazon S3ETL Pipelines

Visualization & Analysis

Power BIExcelCanva Exploratory Data AnalysisStatistical Data AnalysisStar Schema Modeling

Governance & Compliance

HIPAAFERPAPII Handling Data GovernanceData ValidationADA Compliance

Applied ML & AI

Scikit-learntidymodelsOpenAI API Function CallingGrounding / RAGStreamlit

Critical ThinkingCross-functional Collaboration Interpersonal SkillsAdaptability Teamwork

How I work

Validate before I visualize

Every dashboard I ship has been through a data quality and validation pass first. A clean-looking chart built on unvalidated data is worse than no chart at all.

Ground before I generate

Whether it's a compliance report or an AI agent's answer, I don't let output get ahead of its source. My agents cite what they used; my reports document where the numbers came from.

Build for the person who has to trust it

A school administrator, a healthcare client, a student deciding on a semester abroad: the end user isn't a data person. If they can't act on it, the analysis isn't finished yet.

Code samples

Real snippets pulled straight from my own scripts, not illustrative pseudocode.

Census & Election analysis (R) · permutation test

# Two-sample t-test on observed groups
ttest_result <- t.test(formula = bach_higher ~ party_clean,
      data = trial_data, alternative = "two.sided")

n_permutations <- 5000
permutation_statistics <- vector(length = n_permutations)

for(p in 1:n_permutations) {
  permutation_statistics[p] <- t.test(formula = bach_higher ~ party_clean,
      data = trial_data |> mutate(party_clean = sample(party_clean))) |>
      broom::tidy() |> pull(statistic)
}

Census & Election analysis (R) · bootstrap CI

income_data <- state_census |>
  filter(!is.na(median_income)) |>
  pull(median_income)

B <- 10000
boot_medians <- replicate(B, {
  sample(income_data, size = length(income_data),
         replace = TRUE) |> median(na.rm = TRUE)
})

boot_se <- sd(boot_medians)
boot_ci <- quantile(boot_medians,
      probs = c(0.025, 0.975))

Family Involvement & Grades (R / tidymodels) · model workflow

multi_spec <- multinom_reg(mode = "classification") %>%
  set_engine("nnet")

multi_wf <- workflow() %>%
  add_recipe(grade_rec) %>%
  add_model(multi_spec)

set.seed(123)
multi_res <- multi_wf %>%
  fit_resamples(
    resamples = cv_folds,
    metrics = metric_set(accuracy)
  )

STA 631 Modeling Portfolio (R / tidymodels) · LDA spec

lda_spec <- discrim_linear(mode = "classification") %>%
  set_engine("MASS")

lda_wf <- workflow() %>%
  add_model(lda_spec) %>%
  add_recipe(class_rec)

set.seed(123)
lda_cv <- fit_resamples(
  lda_wf, resamples = class_folds,
  control = control_resamples(save_pred = TRUE)
)

Experience

Full role-by-role detail lives in my resume, which I tailor per application. Here's the shape of it:

Aug 2025–Present

Data Analyst Graduate AssistantGVSU Charter Schools Office

Building the ETL & BI platform for 78 schools (see project above)

Mar 2025–Aug 2025

Study Abroad Peer AdvisorGrand Valley State University

Advised 100+ students; organized 17 years of program data

Dec 2023–Dec 2024

Associate Data EngineerCedar Gate Technologies

SQL/Redshift healthcare reporting under HIPAA

Aug 2023–Nov 2023

Data Analyst InternAyata Incorporation Pvt. Ltd

EDA & Power BI dashboards for stakeholder reporting

Education

DEC 2026

M.S. Data Science and Analytics

Grand Valley State University, Allendale, MI · GPA 3.9/4.0

AUG 2023

B.E. Computer Engineering

Tribhuvan University, Institute of Engineering, Kathmandu, Nepal

Let's talk

Open to full-time Data Analyst, Data Engineer, and Data Scientist roles and related roles, available starting December 2026. Open to relocation. Reach out any time.

Emailkarkibi@mail.gvsu.edu

Phone(734) 510-9027

LinkedInlinkedin.com/in/binju-karki

GitHubgithub.com/Binju-16