Open to Data Science & ML roles

Portfolio — AI & Data Science

Umang Panchal

M.Tech AI & Data Science Data Scientist Machine Learning Artificial Intelligence

I build intelligent systems that turn data into decisions.

I design, build and deploy data-driven applications — from an AI business-intelligence platform and an automated data science engine running in production, to privacy-preserving federated learning research for healthcare prediction.

View projects Let's connect
Email Phone GitHub LinkedIn Resume

8.80

CGPA — M.Tech, AI & Data Science

9.20

CGPA — B.Tech, Information Technology

2

Data internships completed

2

Platforms deployed and live

01 About

Working where data engineering meets applied AI.

I work at the intersection of machine learning, data analytics and applied AI — building systems that move from raw, messy data to useful insight and, eventually, to an automated decision. Most of what I build ends up deployed, not left in a notebook.

I'm currently completing an M.Tech in Artificial Intelligence & Data Science at Parul University, Vadodara, after a B.Tech in Information Technology at Silver Oak University. Alongside coursework, I've shipped two production web platforms — one for AI-assisted business intelligence, one for automated data science — and I'm running research into privacy-preserving federated learning for healthcare prediction.

The engineering half matters to me as much as the modelling. Cloud databases, environment-based configuration, serverless constraints, CI/CD — a model is only useful once somebody else can reach it, and keeping it reachable is part of the job.

Focus

Applied ML, data platforms & privacy-preserving AI

Currently

M.Tech AI & Data Science
Parul University, Vadodara

Shipped

2 production platforms
deployed on cloud infrastructure

Researching

Federated learning with
adaptive differential privacy

How I work/ pipeline

Dataingest · clean · validate
AnalysisEDA · statistics · profiling
Machine Learningtrain · tune · evaluate
Intelligenceinsight · explanation · XAI
Decisiondeployed · usable · in production

02 Skills

The stack behind the systems.

Grouped the way I actually use them — from writing the query to shipping the model.

Programming

05
PythonSQLJavaScript HTMLCSS

Data Science

06
PandasNumPyScikit-learn MatplotlibSeabornSciPy

Machine Learning

07
RegressionClassificationClustering Feature EngineeringModel Evaluation Predictive AnalyticsAnomaly Detection

Deep Learning

07
CNNRNNLSTMBiLSTM AttentionTensorFlowPyTorch

AI & Research

06
Federated LearningDifferential Privacy RDPFedAvgExplainable AISHAP

Data & BI

05
Power BIData VisualizationEDA Statistical AnalysisBusiness Intelligence

Backend & Cloud

07
FastAPIMySQLSQLAlchemy AivenVercelGitGitHub

03 Experience

Where I've worked with data.

Internship

Mar 2024 — Jun 2024

Data Science Intern

SAP & Edunet

  • Applied Python to data analysis, statistical modeling and machine learning tasks.
  • Used NumPy, Pandas, Scikit-learn and TensorFlow to process data and build ML models.
PythonNumPyPandas Scikit-learnTensorFlow
Internship

Dec 2023 — Jan 2024

Data Visualization Intern

IBM SkillsBuild

  • Built dynamic, interactive dashboards and reports using Tableau, Power BI and Excel.
  • Applied data visualization techniques to present data in a clear, structured format.
Power BITableauExcel Dashboarding

04 Projects

Three systems, one trajectory.

Each project moves one step further up the stack — from analytics, to automated machine learning, to privacy-preserving AI research.

01 AI · Business Intelligence Live

BI Copilot

AI-Powered Business Intelligence Platform

A full-stack business intelligence copilot. Users upload a business dataset and the platform runs automated analysis, visualization, prediction and anomaly detection, with an AI assistant layered on top for natural-language questions about the data.

  • Automated analysis of CSV / XLSX business datasets
  • AI copilot for natural-language data assistance
  • Predictive analytics, anomaly detection & data-quality profiling
  • Deployed on Vercel with Aiven Cloud MySQL and a custom domain
PythonFastAPIPandas SQLAlchemyMySQLAiven Vercel
Live demo
02 Automated Data Science Live

Data Scientist Agent 2.0

AI-Powered Automated Data Analysis & ML Platform

A production-deployed automated data science platform. Upload a CSV and it profiles the dataset, runs exploratory analysis and statistics, generates visualizations, and drives machine learning workflows end to end — all through the browser, with analysis history persisted.

  • Automated profiling: shape, dtypes, missing values, dataset statistics
  • Automated EDA and visualization with Matplotlib & Seaborn
  • Model building and evaluation workflows
  • Re-architected storage to survive a serverless filesystem — see case study
PythonFastAPIScikit-learn SeabornJinja2MySQL AivenVercel
Live demo
03 M.Tech Research · Privacy-Preserving AI Research

Privacy-Preserving Federated Learning

for Healthcare Prediction · Adaptive Differential Privacy

A federated learning platform where multiple simulated healthcare institutions collaboratively train a shared model without exchanging raw patient records. Adaptive differential privacy adjusts gradient clipping and noise during training to balance model utility against privacy cost under heterogeneous, non-IID data.

  • FedAvg aggregation across simulated healthcare clients
  • Adaptive gradient clipping and adaptive noise multiplier
  • RDP-based privacy accounting, tracking ε at a given δ
  • Explainable AI with SHAP and a four-way experimental comparison
Federated LearningDifferential Privacy RDPFedAvgSHAP Scikit-learnSciPy
Read the research

Earlier work

  • Credit Card Fraud Detection — classification on a highly imbalanced financial transaction dataset in Python.
  • Exploratory Data Analysis of Car Features — EDA on a Kaggle dataset, surfacing relationships between specification, pricing and category.

Code

  • Repositories and notebooks are on GitHub.
github.com/Umangkp1

05 Research / M.Tech Work

Privacy-Preserving Federated Learning for Healthcare Prediction

Training a useful model on sensitive data you are never allowed to centralise — and measuring exactly what privacy costs you.

Adaptive Differential Privacy in a federated setting.

Healthcare data is exactly the kind of data that benefits most from machine learning and is least likely to be shared. Federated learning removes the need to pool it — but model updates themselves can leak information. This work adds differential privacy to federated training, and makes the privacy mechanism adaptive rather than fixed, so the noise budget responds to training rather than being set once and forgotten.

Problem

Institutions holding patient data cannot pool it, and static differentially private training tends to over-noise early rounds and under-protect later ones.

Approach

Federated training with FedAvg, per-client gradient clipping and Gaussian noise, where the clipping threshold and noise multiplier adapt across rounds.

Accounting

Rényi Differential Privacy accounting tracks the cumulative privacy budget, reporting ε at a chosen δ rather than treating privacy as qualitative.

Heterogeneity

Clients are partitioned both IID and non-IID via a Dirichlet distribution, so the method is evaluated under realistic data skew.

Federated training architecture Simulated healthcare clients · conceptual

LOCAL / PRIVATE PRIVACY LAYER SERVER EVALUATION RAW PATIENT DATA STAYS INSIDE THIS BOUNDARY Healthcare Client 1 local dataset · non-IID Healthcare Client 2 local dataset · non-IID Healthcare Client 3 local dataset · non-IID Healthcare Client 4 local dataset · non-IID Healthcare Client 5 local dataset · non-IID Local Model Training no raw data leaves Gradient Clipping Adaptive Control Gaussian Noise RDP accounting · ε at δ FedAvg Aggregation Server receives private updates only weighted parameter average private updates Global Model broadcast next round Performance accuracy · precision · recall F1-score · ROC-AUC Privacy Budget ε tracked via RDP at fixed δ Explainability SHAP feature attribution

Privacy mechanism

per client, per round

01Local gradientcomputed on private data
02Gradient clippingbounded sensitivity
03Adaptive noisemultiplier tuned per round
04Private updatesafe to transmit
05AggregationFedAvg on the server

Raw patient data never leaves the local client. Only clipped, noised parameter updates are shared.

Experimental comparison

four training regimes evaluated

Centralized ML

all data pooled · baseline

Utility

Privacy

Standard Federated Learning

FedAvg · no formal privacy guarantee

Utility

Privacy

Static DP-FL

fixed clipping & noise

Utility

Privacy

Adaptive DP-FL — THIS WORK

adaptive clipping & noise · RDP accounted

Utility

Privacy

The meters above are a conceptual illustration of the privacy–utility trade-off each regime targets, not measured experimental results. Quantitative accuracy, precision, recall, F1, ROC-AUC and ε values are produced by the experiments themselves.

Research technologies

PythonPandasNumPy Scikit-learnSciPyMatplotlib SHAPFederated Learning Differential PrivacyRDP FedAvgExplainable AI

Research areas

Privacy-Preserving MLHealthcare AI Distributed Machine LearningNon-IID Learning Privacy–Utility Trade-offExplainable AI

06 Education

Academic record.

M.Tech — Artificial Intelligence & Data Science

May 2025 — Present

Parul University, Vadodara

8.80CGPA

Currently pursuing. Research work: privacy-preserving federated learning with adaptive differential privacy for healthcare prediction.

B.Tech — Information Technology

May 2021 — May 2025

Silver Oak University, Ahmedabad

9.20CGPA

Coursework: Machine Learning, Deep Learning, Analysis of Algorithms, Engineering Mathematics, DBMS, Python, OOPS.

Achievements

  • 2nd rank in the university, 2nd year — Silver Oak University
  • Project selected for Techstars Global Startup Weekend, PDEU

Certifications

  • Advanced Data Science & AI/ML Program — Learnbay, Bangalore (IBM & Microsoft certification program)

07 Contact

Let's build something intelligent.

Open to data science, machine learning and AI engineering roles, research collaborations and interesting problems. Email is the fastest way to reach me.

Email umangpanchal1818@gmail.com Write to me → Phone +91 91736 80818 Call → LinkedIn umang-panchal Connect → GitHub Umangkp1 View code →
Send an email See a live system