Data Science · AI Engineering · Deep Learning

AI built with real-world constraints in mind.

AI engineer with a Master’s in Artificial Intelligence, focused on building practical LLM systems with LangChain, multi-agent workflows, and deep learning. I enjoy working end-to-end, from data and training to deployment, and keeping things reliable and easy to understand.

View projects Email me

Skills & tools

Skills

Everything I’ve used recently...

Python PyTorch TensorFlow NLP LangChain Pandas Java C++ Optimization Algorithms (PuLP) Reinforcement Learning VAEs CNNs Transformers Computer Vision AWS (S3, EC2, Athena) Azure Databricks SQL Tableau Data Validation/ QA BeautifulSoup MATLAB GitHub Git

Project slideshow

Projects

Full-Stack · FastAPI · React · ML Pipelines

Deepfake Detection Tool & Chrome Extension

Full-stack deepfake platform with a Chrome extension, React dashboard, FastAPI services, and a ClearML-orchestrated ML pipeline on remote GPUs.

Chrome extension scans webpage images and overlays deepfake-probability scores in real time.
ClearML pipeline for dataset upload, preprocessing, training/testing, and experiment tracking; React frontend + FastAPI APIs with docs, health checks, and browser-side inference.
Tools: React, FastAPI, ClearML, Poetry, GitHub Actions, Chrome extension, Colab GPU agents.

GitHub repo

Multi-Agent · LangChain · RAG

Multi-Agent Finance Risk Analysis

LangChain and RAG-powered finance system with specialized agents that pull real-time data, debate, and converge on risk-aware investment calls, inspired by BlackRock.

Six-agent workflow handling data retrieval, fundamentals, sentiment, valuation, risk, and report writing.
Round-robin debate manager to surface disagreements and drive consensus on investment recommendations.
Supports local/Ollama or API-backed models; includes reproducible setup, tests, and sample outputs.

GitHub repo

Deep Learning · Medical Imaging

Dementia classification using MRI scans

Ensembled ResNet-based models to classify dementia severity with tight GPU limits, keeping inference fast enough for interactive review.

PyTorch preprocessing and augmentation tailored to noisy MRI data.
Handled class imbalance with resampling strategies to lift minority-class accuracy.
Packaged outputs into a lightweight review flow for clinicians.

Kaggle notebook

Generative Models · Data Quality

Variational Autoencoder oversampling

Used VAEs to synthesize training samples for skewed datasets, improving model robustness without expensive data collection.

Explored latent spaces to monitor mode coverage and reconstruction quality.
Compared VAE-driven oversampling to classical SMOTE variants on tabular data.
Wrapped training into a reproducible Colab workflow for quick tuning.

Open in Colab