Open to Data Science and Applied ML internships

JUNIOR DATA SCIENTIST · APPLIED ML · AI SYSTEMS

I turn large, messy datasets into tested decisions.

I build large-scale analytics, statistical experiments, human-reviewed ML workflows and grounded retrieval systems - with measurable evaluation, public code and deployable interfaces.

View projectsResume
15.95Mcomplaint records processed
+0.769 ppobserved A/B conversion uplift
Human gatedeterministic-first ML auditing
0.94729-case strict hybrid RAGAS composite
OPTIONAL 3D SYSTEM MAPLOAD ON REQUEST
Three project systems shown as a static fallbackLoading 3D engine…

Use left and right arrow keys to rotate the 3D scene, up and down arrows to tilt it, and Home to reset the view.

SYSTEM 01 · LARGE-SCALE DATA SCIENCE Turn 15.95M raw complaints into decision-ready evidence.

Chunked processing, Parquet, risk analysis, NLP and holdout forecasting converge into an executive decision layer.

Three flagship systems show Data Science, ML reliability and retrieval engineering; one focused case study shows statistical decision-making.

Four projects. One evidence-first profile.

Three flagship systems show scale, ML reliability and retrieval engineering; one focused experiment shows statistical decision-making.

01 / LARGE-SCALE DATA SCIENCE

Customer Complaint Intelligence Platform

Probleman 8-9 GB complaint file is too large and too messy for dashboard-time analysis.
System decisionchunked processing, Parquet summaries and separate NLP training.
What it enablesdecision-ready analytics with 3.57% six-month holdout MAPE.

A large-scale Data Science platform that converts an 8-9 GB CFPB complaint file into reusable Parquet analytics, company risk signals, NLP routing, validated forecasts and executive recommendations.

15.95Mcomplaints
3.57%forecast MAPE
14core tests passing
PythonPandas / PyArrowProphetscikit-learnStreamlit
Engineering decisions, role and limitations
My role
Built the preprocessing pipeline, Parquet outputs, analytics modules, risk logic, forecasting, NLP models and Streamlit dashboard.
Key decision
Pre-aggregate expensive analysis and read Parquet summaries at runtime instead of loading the 8-9 GB CSV in the app.
Hardest problem
Balancing scale, business actionability and honest interpretation of raw complaint volume.
Boundary
Complaint volume is not normalized by customer base; narrative coverage is 23.84%; product and issue classifiers are documented at 75.28% and 62.39% accuracy rather than overstated.
GitHub Video Docker image Evaluation methodology
INTERACTIVE 3D · DATA SCIENCE STORYRAW CSV ONLINE
Customer Complaint Intelligence pipeline posterPreparing 3D scene…

Use left and right arrow keys to rotate the 3D scene, up and down arrows to tilt it, and Home to reset the view.

00:00 / 00:56
8–9 GB raw complaint sourceThe pipeline starts with a file too large to load casually inside an interactive dashboard.

Why it matters: Scale is an engineering constraint, so the animation begins with the memory problem.

REPOSITORY EVIDENCE

Selected current repository assets - shown at natural aspect ratio and available full size.

Executive dashboardOpen full size
System architectureOpen full size
Six-month forecastingOpen full size
NLP complaint intelligenceOpen full size
02 / STATISTICAL EXPERIMENTATION

FOCUSED DATA SCIENCE CASE STUDY

Marketing A/B Testing and Statistical Experiment Analysis

Did the advertisement group genuinely achieve a higher conversion rate than the PSA control?

An end-to-end statistical experiment analysis across 588,101 users, separating observed uplift, statistical reliability, effect size and business rollout requirements.

588,101users analysed
+0.769 ppabsolute uplift
95% CI+0.595 to +0.943 pp
~130users per additional conversion
Decision Positive statistical evidence; recommend a controlled, cost-aware rollout or follow-up experiment.

The two-sided p-value was 1.705 × 10⁻¹³, but the decision is framed around effect size, uncertainty and business economics. Causal interpretation still depends on valid assignment and logging.

PythonPandasSciPyStatsmodelsMatplotlib
GitHub ↗ Experiment report ↗
Observed conversion liftPSA 1.785% → Ad 2.555%
Uncertainty checkSimulated uplift remains well above zero
03 / ML RELIABILITY & HUMAN REVIEW

Agentic ML Audit Copilot

Problemrisky datasets can make weak models look trustworthy.
System decisiondeterministic audits pause at a human review gate.
What it enablesautomated pytest coverage, baseline evidence, MLflow and SHAP reporting.

A deterministic-first pre-training audit system that surfaces data-quality, leakage and imbalance risks, pauses risky workflows for human review, then records baseline, MLflow and explainability evidence.

HITLdecision gate
Deterministicrisk checks
pytestautomated coverage
Pythonscikit-learnLangGraphMLflow / SHAPFastAPI
Engineering decisions, role and limitations
My role
Designed the workflow, audit modules, risk routing, API, Streamlit interface, MLflow tracking and report flow.
Key decision
Python remains the judge; the LLM is limited to explanation, report writing and audit Q&A.
Hardest problem
Making leakage, imbalance and data-quality findings visible without pretending the system is AutoML.
Boundary
Single-machine tabular classification/regression review, not enterprise governance or automated model selection.
GitHub Live app Watch walkthrough Docker image ↗
INTERACTIVE 3D · ML RELIABILITY STORYDATASET ARRIVING
Agentic ML Audit Copilot workflow posterPreparing 3D scene…

Use left and right arrow keys to rotate the 3D scene, up and down arrows to tilt it, and Home to reset the view.

deterministic checks risk stops workflow human-approved path
00:00 / 00:52
Dataset enters the auditA CSV and target column are loaded, profiled and converted into a visible matrix before any model is allowed to train.

Why it matters: The workflow starts with evidence about the dataset—not assumptions about the model.

REPOSITORY EVIDENCE

Selected current repository assets - shown at natural aspect ratio and available full size.

Human review gateOpen full size
Executive audit dashboardOpen full size
Human-in-the-loop API workflowOpen full size
04 / APPLIED AI & RETRIEVAL

Enterprise RAG Assistant

ProblemRAG is unreliable for exact numerical questions and weak at exact-token retrieval.
System decisionroute analytics to Pandas and semantics to hybrid retrieval.
What it enablesgrounded citations plus a 29-case strict hybrid RAGAS evaluation with 0.947 composite.

A two-route enterprise assistant: exact CSV/Excel analytics go to deterministic Pandas computation; semantic questions use BGE + BM25 hybrid retrieval, CrossEncoder reranking and source-grounded answers, served through FastAPI with a React + Vite frontend.

0.947strict hybrid RAGAS composite
0.966faithfulness
29strict evaluation cases
FastAPIReact + ViteBGE + BM25CrossEncoderPandas
Engineering decisions, role and limitations
My role
Built multi-format ingestion, persistent Chroma storage, hybrid search, CrossEncoder reranking, query routing, Pandas analytics, FastAPI integration and the React + Vite web frontend.
Key decision
Separate computation from retrieval: CSV/Excel aggregations go to Pandas; semantic questions go to RAG.
Hardest problem
Exact IDs and numerical questions looked confident but failed under pure vector search, requiring BM25 and deterministic routing.
Boundary
Current public evidence is the repository, video walkthrough and evaluation report. The separate 1,642-case production benchmark still has an open latency acceptance gate.
GitHub Video Docker profile Evaluation source of truth
INTERACTIVE 3D · RETRIEVAL STORYDOCUMENTS ARRIVING
Enterprise RAG retrieval workflow posterPreparing 3D scene…

Use left and right arrow keys to rotate the 3D scene, up and down arrows to tilt it, and Home to reset the view.

00:00 / 00:58
Documents become searchable evidenceSeven validated formats - PDF, DOCX, CSV, JSON, TXT, XLS and XLSX - enter the ingestion layer with source metadata.

Why it matters: The system has to preserve source identity before retrieval can be trusted.

REPOSITORY EVIDENCE

Selected current repository assets - shown at natural aspect ratio and available full size.

Grounded answer with source evidenceOpen full size
29-case strict hybrid RAGAS evaluationOpen full size
Read-only knowledge registry interfaceOpen full size

TECHNICAL CAPABILITY

Skills grouped by the work they enable.

No percentage bars. The page emphasizes skills demonstrated in projects; Advanced Excel and Power Query are currently being strengthened.

01

Data & SQL

Python, SQL, Pandas, NumPy, PyArrow, Microsoft SQL Server, EDA, validation, chunked processing and Parquet workflows.

02

Statistics & Experimentation

A/B testing, p-values, confidence intervals, chi-square tests, effect sizes, power planning, SciPy and Statsmodels.

03

Machine Learning

scikit-learn pipelines, classification, regression, cross-validation, forecasting, class imbalance, leakage checks, model evaluation and SHAP.

04

NLP & Forecasting

TF-IDF, text classification, topic modeling, Logistic Regression, time-series holdout validation and Prophet forecasting.

05

RAG & Applied AI

Embeddings, BGE, ChromaDB, BM25, hybrid retrieval, query routing, CrossEncoder reranking, grounding and LangGraph.

06

Engineering & MLOps

FastAPI, REST APIs, React + Vite, Streamlit, Docker, MLflow, pytest, Ruff, Git, GitHub Actions, logging and reproducible configuration.

Shivam RajputOPEN TO INTERNSHIPS

ABOUT

Junior Data Scientist with an engineering mindset.

I am pursuing B.Tech CSE with an AI & ML specialization at Uttaranchal University. I enjoy turning messy data and ambiguous questions into analyses and systems whose assumptions, evidence and outputs can be inspected.

My public work spans large-scale complaint analytics, statistical experimentation, ML audit reliability and enterprise RAG - supported by tests, evaluation artifacts, APIs, demos and documented trade-offs.

EducationB.Tech CSE (AI & ML), 2023-2027
CGPA8.7 / 10.0
LocationDehradun, India
Current focusData Science interviews, SQL and Advanced Excel
GitHub LinkedIn LeetCode Kaggle Docker Hub

ONE-PAGE ATS RESUME

The same work, written for recruiters and ATS.

One page, ATS-readable, four selected projects and quantified outcomes without unsupported experience claims.

Download resume

CONTACT

Let's discuss a Data Science or Applied ML opportunity.

Open to internships where I can work on analytics, machine learning, NLP or reliable AI systems. Use the form or email me directly.

shivamrajput.datascientist@gmail.com
Open Gmail compose

Form submission · direct email available above.

Privacy: your name, email and message are used only to respond to this enquiry.