Vision-language models deep dive
Builds a VLM from scratch in PyTorch, then a LLaVA-style model with LoRA, local document field extraction, single-document visual QA, and retrieval-augmented visual QA with ColQwen2.
Projects
Deep dives, GenAI experiments, LLMOps labs and ML foundations, all on GitHub. Where a project has a write-up, it's linked next to the code.
The code behind posts on this site.
Builds a VLM from scratch in PyTorch, then a LLaVA-style model with LoRA, local document field extraction, single-document visual QA, and retrieval-augmented visual QA with ColQwen2.

A full implicit-feedback pipeline on MovieLens 25M and H&M: baselines, ALS and BPR, FAISS two-stage retrieval with LightGBM ranking, then external validation with RecBole.
Three specialised brains: Strategist for NL-to-SQL analytics with Plotly charts, Resolver for operational RAG, and Navigator for documentation RAG that says "I don't know".

Ingests RFP and contract PDFs into Chroma with Hugging Face embeddings, with retrieval and Q&A through Groq from a CLI or a Streamlit UI.
Fine-tuning, agentic workflows, tool calling and retrieval.
Fine-tunes Llama 3.2 3B Instruct on DeepSeek-R1 chain-of-thought data (~130K rows) with QLoRA, early stopping and validation monitoring.
A LangGraph workflow with separate title and content nodes that writes SEO-friendly posts on any topic, with model choice and Markdown export.
A persona-driven chat assistant that switches between Groq-hosted open models and OpenAI's GPT-5 family, adapting request parameters per provider.
A customer-support agent that calls backend tools over JSON-RPC, packaged with Docker and a Kubernetes deployment.
Contract PDFs into a persistent Chroma store with source-aware retrieval and RAG answers, from a CLI or a Streamlit UI.
LLM apps taken all the way to deployment, CI/CD and monitoring.
A RAG chatbot over medical PDFs with a Flask app, Docker image and a Jenkins pipeline.
Itinerary planning with LangChain and Groq, deployed on Kubernetes with ELK-stack logging.
Recommends products from titles and reviews using AstraDB vector search, monitored with Prometheus and Grafana.
An LLM recommender over anime metadata with Chroma retrieval, packaged for Kubernetes.
Generates study quizzes with LLMs, deployed through a GitOps workflow on Kubernetes.
Classic data science projects from my postgraduate studies.
Predicts property values for a market-entry decision using Ridge and Lasso regression.
Scores online-education leads by likelihood to convert with logistic regression.
Clusters countries on socio-economic indicators to direct development funding.
Models demand for shared bikes with multiple linear regression.
Exploratory analysis of loan data to find who is most likely to default.
Insights from a century of movie data, from 1916 to 2016.
Experiments, exercises and work in progress live on my profile.