Skip to content

Naram Charan

◢SPECIALIZING INComputer Vision

“I teach machines to predict things.”

View Projects→Project Aegis3D→
J.A.R.V.I.S // MARK XLII
01PORTFOLIO

Selected Projects

Medical imaging, forecasting, and the pipelines behind them. Explore the results, my contribution, and what I learned.

MK-04Medical Imaging · Deployed
Featured projectLIVE

RSNA Pneumonia Detection

PRIMARY METRIC

83% recall

Pneumonia screening on chest radiographs — transfer learning across three CNNs on the RSNA challenge dataset, then shipped: DICOM pipeline, FastAPI service, React interface, one Docker container live on Azure.

  • 83% pneumonia recall · 0.79 F1 on 1,836 held-out studies
  • Found RandomCrop discarding 95% of every 1024² X-ray — the single biggest fix
  • Live on Azure: one container serves API + UI at $5.06/month
  • PyTorch
  • pydicom
  • FastAPI
  • React
  • Docker
  • Azure
Live Demo ↗View Code ↗
MK-04 / PREVIEWApplication capture
RSNA application displaying the real radiograph for stored study Case_1
Live application · stored study shown · 12 October 2026View source ↗
MK-03Time-Series Forecasting
Featured projectCOMPLETE

Walmart Store Weekly Sales Forecasting

PRIMARY METRIC

0.9555 validation R²

Global multi-series forecasting across ~3,000 Walmart store-department pairs — one model learns them all, driven by a custom recursive walk-forward engine that feeds each week's predictions back in as the next week's lag features.

  • Validation R² 0.9555 — LightGBM beat XGBoost & Random Forest
  • Recursive walk-forward engine built from scratch, no forecasting libs
  • Caught & fixed temporal leakage masquerading as R² ≈ 0.98
  • Python
  • LightGBM
  • XGBoost
  • Scikit-Learn
  • Pandas
View Code ↗
MK-03 / PREVIEWPublished results
Published validation comparison of Random Forest, XGBoost and LightGBM for Walmart sales forecasting
Published validation R² · bars start at zeroView source ↗
MK-02 / PREVIEWPublished results
Held-out test F1 comparison: Random Forest 91.33%, XGBoost 94.74%, neural network 96.23%
Test-set F1 · values from the project READMEView source ↗
MK-02Classification
COMPLETE

E-Commerce Customer Churn Prediction

PRIMARY METRIC

98.28% acc

Four-model churn system — Logistic Regression, Random Forest, XGBoost and a 4-hidden-layer PyTorch network. The deep net tops the leaderboard; XGBoost is selected anyway, because a half-point of accuracy doesn't buy a GPU bill.

  • 98.28% accuracy · 94.74% F1 on the held-out test set
  • PyTorch net hit 98.76% — rejected: marginal gain, heavy compute
  • Optuna over 10-fold Stratified CV, scored on recall
  • XGBoost
  • PyTorch
  • Scikit-Learn
  • Optuna
View Code ↗
MK-05 / PREVIEWArchitecture illustration

Decoder-only Transformer

  1. 01Character input
  2. 02Embedding + position
  3. 034 decoder blocks
  4. 04Next-character logits

4 attention heads · 64-dim embeddings · 32-character context

Project architecture · illustration, not execution outputView source ↗
MK-05Learning Project · NLP
ONGOING

Handbuilt Transformer

PRIMARY METRIC

1.82 val loss

A mini GPT and a mini GPT tokenizer, both built from scratch in PyTorch while learning how ChatGPT-style models work: a small decoder-only Transformer that writes Shakespeare-like text, and a GPT-2-style byte-level BPE tokenizer. A learning project I'm still extending.

  • ~0.21M-param decoder-only GPT: loss 4.41 → 1.66 train, 1.82 validation
  • GPT-2-style BPE tokenizer: regex chunking + byte-level merges, lossless round-trip
  • Learning project, still growing — next up: train the GPT on BPE tokens
  • PyTorch
  • Transformers
  • BPE Tokenizer
  • NLP
View Code ↗
MK-01 / PREVIEWArchitecture illustration

Book-data pipeline

  1. 01Book pages
  2. 02Extract + validate
  3. 03SQLite records
  4. 04CSV export

Title · price · rating · availability · description · URL

Project architecture · illustration, not execution outputView source ↗
MK-01Data Engineering
COMPLETE

Book Data Scraping & Database Pipeline

PRIMARY METRIC

980+ in <30 min

Automated multi-page scraper using BeautifulSoup + SQLAlchemy ORM, extracting structured data (title, price, rating, availability, description, URL) into a relational schema.

  • 980+ books scraped in under 30 minutes
  • BeautifulSoup + SQLAlchemy ORM pipeline
  • Relational SQLite schema + clean CSV exports
  • BeautifulSoup
  • SQLAlchemy
  • SQLite
  • Pandas
View Code ↗
02PORTFOLIO

Skills & Systems

Tools I use, with evidence from the projects where I applied them.

◢ SYS-01 · NEURALONLINE

Deep Learning & GenAI

  • PyTorch
  • Transfer Learning · CNNs
  • Transformers · Self-Attention
  • BPE Tokenization
  • Prompt Engineering

MK-04 & MK-05 — ResNet-34 transfer learning on chest X-rays at 83% pneumonia recall; causal self-attention and a BPE tokenizer written by hand in PyTorch.

◢ SYS-02 · MODELSONLINE

Machine Learning

  • XGBoost
  • LightGBM
  • Random Forest
  • Logistic Regression
  • Scikit-Learn
  • Optuna

MK-03 & MK-02 — LightGBM at 95.55% R² across ~3,000 series; XGBoost at 98.28% accuracy, tuned with Optuna.

◢ SYS-03 · SIGNALSONLINE

Data Intelligence

  • Pandas
  • NumPy
  • Feature Engineering
  • KNNImputer
  • Imbalance Handling
  • RobustScaler

MK-03 — recursive walk-forward feature engine; diagnosed and killed the leakage inflating R² to 0.98.

◢ SYS-04 · COREONLINE

Engineering Core

  • Python 3 · Advanced OOP
  • Data Structures
  • FastAPI · REST APIs
  • SQL · Schema Design
  • SQLAlchemy
  • Docker · Azure
  • Claude Code · Agentic Dev
  • Git Workflow
  • BeautifulSoup

MK-04 — FastAPI serving a model and a React bundle from one container on Azure. This site itself — Next.js + WebGL, built by directing Claude Code.

03PORTFOLIO

Origin Story

It started with a man in a suit of armor — the coolest hero I'd ever seen. But my interest went past the armor, to the AI he talked to. Because you know what was even cooler than the suit? His AI.

I'm a 3rd-year CS student specializing in AI & Machine Learning, focused on end-to-end data pipelines, predictive modeling, and scalable systems — not notebook exercises.

I train deep networks on medical images, build transformers from scratch to understand how they work, tune tree-based classifiers with hyperparameter optimization, and obsess over data-separation methodology so models hold up in production, not just on a leaderboard.

EDUCATION

B.Tech Computer Science Engineering

Artificial Intelligence & Machine Learning

GD Goenka University, Gurugram · 3rd Year

CGPA AVERAGE
8.98 / 10.0
GRADUATION
Expected May 2028
CERTIFICATIONS

Deep Learning Specialization

DeepLearning.AI

Jul 2026 · VERIFY →

Claude Code: A Highly Agentic Coding Assistant

DeepLearning.AI · Anthropic

Jul 2026 · VERIFY →

Machine Learning Specialization

DeepLearning.AI · Stanford University

May 2026 · VERIFY →

Databases & SQL for Data Science with Python

IBM · Coursera

Nov 2025 · VERIFY →

Python for Everybody Specialization

University of Michigan · Coursera

Mar 2025 · VERIFY →

Prompt Engineering & Generative AI

Google · Vanderbilt University

Mar 2025 · VERIFY →

04PORTFOLIO

How I Think

The whole field on one branch diagram — split by the question that decides your entire approach: is the data a table, or isn't it. Open any model to see what I built with it.

PROJECTION · ML TAXONOMY8/11 MAPPED TO PROJECTS
ROOT

MACHINE LEARNING

WHAT SHAPE IS THE DATA?

TABULAR DATA

TRADITIONAL ML

Rows and columns, engineered features. The model learns from structure you designed.

SUPERVISED

Labelled targets — the answer exists in the training data.

UNSUPERVISED

No labels — the model finds the structure itself.

UNSTRUCTURED DATA

DEEP LEARNING

Pixels, tokens, signals. The model learns the features instead of being handed them.

COMPUTER VISION · CNN

Convolutions learn spatial features — edges, textures, structure.

SEQUENTIAL MODELS

Order carries meaning — state is passed along the sequence. Studied in theory: these are the building blocks transformers were built to replace, and knowing exactly where they break is what makes attention make sense.

TRANSFORMERS

Attention replaces recurrence — every position sees every other in one step. This is the architecture running inside ChatGPT, Gemini and Claude, and the reason the modern era of AI happened when it did.

HOW TO READ THIS. A filled cyan dot means I've shipped a project with that model — open the node to jump to the code. Amber is in progress. FOUNDATION means I've studied it in theory and understand how it works, and it carries no project by design: RNN → LSTM → attention is the line the modern era of AI was built on, and the architecture it ends at is the one running inside ChatGPT, Gemini and Claude today. A hollow dot is a method I know but haven't built with yet.

05PORTFOLIO

Contact

Channels are open. Reach out for internships, collaborations, or to talk shop about ML.

jarvis@naramcharan:~$ contact

$ whoami

Naramreddy Charan Kumar Reddy — AI & Machine Learning Engineer

$ list --channels

Open in Gmail ↗

$

DESIGNED & BUILT BY NARAM CHARAN · POWERED BY J.A.R.V.I.S

Helmet model: “Ironman Mark III Helmet *free*” by Demonic Arts, licensed CC BY 4.0; textures resized and compressed.