/01Scientific ML · Drug/Dye Design
NIR Fluorophore Designer
An ML tool that predicts the optical properties of near-infrared fluorophores — dyes that
emit above 600 nm, the most useful region for deep-tissue bioimaging — and generates
novel candidate molecules. Paste a SMILES string and a solvent to predict absorption,
emission, quantum yield, and molar absorptivity; or set an objective and let it design and
rank brand-new NIR dyes. An educational reimplementation inspired by the FLAME framework
(Nature Communications, 2025), focused on the long-wavelength subset, with a
gradient-boosting vs. GNN model comparison.
RDKitLightGBMGradient Boosting
GNN / MPNNCheminformaticsStreamlit
/02Research · Healthcare AI
Aphasia Detection & Severity Prediction
End-to-end ML pipeline that detects aphasia and classifies its severity from clinical
speech transcripts. Built on acoustic and linguistic features (MFCCs, pitch, jitter,
formants, speech rate) from the AphasiaBank corpus. Three architectures trained and
compared — CNN, GRU, and a hybrid CNN+GRU. Developed during a research internship at IIT.
TensorFlowLibrosaPraat-Parselmouth
CNNGRU
/03GenAI · Multi-agent NLP
Homework Helper
A multi-agent AI app that solves homework questions step by step. It analyzes intent and
difficulty with NLP, then routes the query through a pipeline of agents — clarification,
solution, QA review, and a concise answer — before returning a clean, verified response.
StreamlitspaCy / TransformersGroqAgentic pipeline
/04GenAI · RAG
Cold Mail Generator
Paste a company's careers-page URL and get a personalized cold email instantly. It scrapes
the listing, extracts role and skills as structured JSON with an LLM, matches your portfolio
from a vector store by required skills, and drafts a tailored email.
LLaMA 3.3 70BLangChainChromaDBGroqStreamlit
/05Computer Vision
Sign Language Detection
Real-time ASL recognizer for A–Z letters and 0–9 digits via webcam. MediaPipe extracts 21
hand landmarks per frame, a Random Forest classifier predicts the sign, and the result is
overlaid live on the video feed.
MediaPipeOpenCVScikit-learnRandom Forest
/06Recommender Systems
Movie Recommender System
A content-based recommender built from scratch — give it a movie and it returns five with a
similar vibe. Each title's genre and plot become TF-IDF vectors, and cosine similarity
surfaces the closest matches across a 10,000-movie TMDB dataset.
Scikit-learnTF-IDFCosine similarityPandas
