I build AI/ML systems
and the backends behind them.

A kiosk for small clinics: a patient uploads a chest CT scan and gets an AI read of possible lung nodules, even when no radiologist is around.

I built the backend for a lung nodule detection system for rural clinics, where patients upload CT scans and get an AI analyzed report with no radiologist on site.

• I served a 0.32 MB INT8 quantized MobileUNet3D over a REST API, so it runs fully on device without a GPU.
• I used FastAPI with Redis and RQ to queue scans asynchronously, and a LangGraph ReAct agent for clinical reasoning.
• I added Gemini 2.0 Flash Live voice in 10+ Indian languages, FHIR R4 report export and Supabase row level security.

• 0.32 MB INT8 model
• 120x smaller than teacher
• Dice 0.803
• AUC 0.800

• Python
• PyTorch
• FastAPI
• Redis
• LangGraph
• Gemini
• Supabase

A hiring assistant: it reads a stack of resumes, ranks the ones worth interviewing, writes interview questions and reads how calm or stressed the candidate sounds.

I built a recruiter facing system that shortlists resumes, generates interview questions and analyses candidate emotion, each part running as its own service.

• I tested the ranking model on 250 resumes across 5 roles against human recruiter judgments and got 86.4% Precision@10.
• Throughput came out at 50 resumes per 3.2 minutes, with about 2.5 seconds to generate a question.
• I built the ranking engine on Mistral 8x7B and Sentence Transformers, and stored sessions and video in Supabase with JWT auth.

• 86.4% Precision@10
• 250 resumes across 5 roles
• 50 resumes per 3.2 min

• Python
• FastAPI
• Mistral 8x7B
• Sentence Transformers
• Supabase
• React

A tutor that notices when a student is confused or bored from their words, voice and face, then eases up or pushes harder on the next question.

I built a tutoring system that reads student emotion from text, voice and webcam in real time and adapts lesson content and quiz difficulty as it goes.

• I deployed a LoRA fine tuned TinyLLaMA as the lesson reasoning backend on a standard 8GB GPU.
• A controlled student pilot returned 85.2% approval on response clarity and tone.
• I indexed OpenStax textbooks with FAISS for grounded lessons, used Whisper for voice and a CNN for facial sentiment.

• 85.2% approval in pilot
• Trained on a standard 8GB GPU

• Python
• TinyLLaMA
• LoRA
• FAISS
• Whisper
• Supabase

A screening tool that estimates how likely someone is to have diabetes from routine health measurements.

I designed a classification pipeline to predict diabetes risk from an imbalanced Kaggle dataset during my internship.

• I benchmarked Extra Trees, XGBoost and SVM under consistent train and test splits, then picked Extra Trees on comparative test accuracy.
• I built the preprocessing layer with median imputation for missing values and correlation based feature pruning.
• I checked pipeline decisions with class distribution and model output plots in Matplotlib.

• Extra Trees selected on test accuracy

• Python
• Scikit learn
• XGBoost
• Pandas
• Matplotlib