Things I’ve built — and shipped.

Computer Vision

Suspicious Crowd Behavior Detection

Final-year project, co-developed for real-world environments. Transfer learning with FAIR's X3D-S network extracts spatio-temporal features and an XGBoost classifier detects anomalies; a full-stack React, FastAPI and MongoDB app operationalizes the pipeline with a multi-class model that categorizes complex human behaviors.

ROC-AUC

0.924

Precision

91.5%

Accuracy

85.2%

Multi-class acc.

86.8%

• PyTorch
• X3D-S
• XGBoost
• FastAPI
• React
• MongoDB

Automation · GenAI & LLMs

AI Job-Hunt Automation Pipeline

An end-to-end automation MVP. Playwright scrapes and normalizes job postings, a matcher ranks them with keyword heuristics plus optional embedding similarity (Gemini or Hugging Face), and a two-step LLM flow extracts the relevant experience and generates a tailored LaTeX résumé for each top match — provider-agnostic across OpenRouter, Groq and SiliconFlow.

• Headless-Chromium scraping with deduplication and a normalized JSON schema
• Hybrid ranking: keyword heuristics plus embedding similarity
• Two-step LLM tailoring that outputs a LaTeX résumé per job

• Python
• Playwright
• LLMs
• Embeddings
• LaTeX

MLOps

End-to-End MLOps Pipeline

DVC versions the data, MLflow tracks experiments and GitHub Actions runs CI/CD. Apache Airflow orchestrates the ETL DAGs, and FastAPI services are deployed on AWS EC2 with Celery workers for background jobs.

• Reproducible pipelines with DVC data versioning and MLflow tracking
• Airflow DAGs orchestrating ETL and retraining
• FastAPI services on AWS EC2 with Celery workers

• AWS
• Airflow
• FastAPI
• Docker
• MLflow
• DVC
• Celery

GenAI & LLMs

Context-Aware Multimodal Recommender

Fuses user preferences with live OpenWeatherMap data through FAISS vector search, and serves dynamic clothing and activity recommendations from transformer models fine-tuned on synthetic data — with millisecond-level latency.

• Hybrid retrieval over preferences and live context with FAISS
• Transformers fine-tuned on synthetic data
• Millisecond-latency recommendation serving

• Python
• RAG
• FAISS
• Transformers
• OpenAI

Healthcare AI

Multimodal Cancer Diagnostic System

A fusion framework that classifies cancer types from clinical, genomic and imaging data, using SHAP and LIME for interpretability and cross-modal attention to correlate unstructured medical text with genomic sequences.

• Clinical, genomic and imaging modalities in one model
• Cross-modal attention between medical text and genomic sequences
• Explainable predictions with SHAP and LIME

• Python
• Explainable AI
• SHAP
• LIME
• Attention