Atanu Adhikari

Student GitHub Reviewer

An AI agent that reviews a student's GitHub portfolio. A LangGraph pipeline drives Llama 3.1 (via Groq), served by a FastAPI backend on Render with a Streamlit UI.

• LangGraph
• Llama 3.1
• Groq
• FastAPI
• Streamlit
• Render

Infant State Recognition from Cry Audio

Classifies infant cry audio into 8 need states: hungry, tired, belly pain, burping, scared, lonely, discomfort and cold/hot. 511 cleaned clips become a 218-dim librosa feature vector (MFCC, delta-MFCC, log-mel, chroma, spectral contrast, tonnetz, ZCR, RMS) fed to a cost-sensitive Random Forest (300 trees, balanced class weights, 5-fold stratified CV).

Test accuracy

81.6%

Macro F1

32.0%

Honest read: the dataset is heavily imbalanced (~75% "hungry"), so accuracy flatters the model. Macro F1 is the metric I'm working to improve. Next: data augmentation and a CNN on log-mel spectrograms.

• Python
• librosa
• scikit-learn
• Random Forest
• NumPy
• Pandas

Flight Delay Risk Prediction

Predicts, the night before departure, whether a New York flight will arrive 15+ minutes late, using 336,776 departures from 2013 (nycflights13). Past-only features with a leakage test, a chronological train/validation/test split, logistic regression vs XGBoost, Platt calibration with rolling recalibration, and a what-if rescheduling analysis.

PR-AUC (XGBoost)

0.468

Prior baseline

0.335

Honest read: December test set. Rolling recalibration cut calibration error (ECE) from 0.102 to 0.047, but at precision ≥ 0.5 recall is only 0.035. There's no weather data, which limits day-ahead accuracy.

• Python
• XGBoost
• scikit-learn
• Pandas
• Jupyter

Cognify: AI Career Guidance Platform

Co-built at Construct Hackathon and launched on Product Hunt. An AI agent asks students cognitive questions to uncover their real interests, then generates personalised career suggestions, roadmaps and college options.

• React
• Vite
• AI agent
• Construct Hackathon
• Product Hunt

MediTrack

A personal health record manager to upload, organise and share medical reports securely. JWT auth, role-based access, protected routes and CRUD REST APIs. Frontend on Vercel, backend on Render.

• React
• Node.js
• Express
• MongoDB
• JWT
• Vercel
• Render

QuickCart

A React e-commerce cart: products fetched from MockAPI, add-to-cart from product cards, quantity controls (1–10), real-time totals and a simulated checkout flow. Multi-page navigation with React Router, responsive on mobile, tablet and desktop. Deployed on Vercel.

• React
• React Router
• JavaScript
• MockAPI
• Vercel

Spendly

An expense tracker built step by step with Claude Code. The Flask app has its landing, register and login pages in place; authentication, profile and add/edit/delete expenses are the next steps.

• Python
• Flask
• pytest
• Claude Code