Writing
Blog
Notes and deep dives on retrieval-augmented generation, LLM fine-tuning, agentic systems, and running GenAI in production.
Sampling Your Way to Generation: Inside the VAE
A VAE's encoder outputs a distribution, not a single code, and that one change is what lets it generate, not just reconstruct. Here's how reconstruction and KL shape the latent space, why the reparameterization trick keeps training differentiable, and how to read blur, posterior collapse, and bad samples as symptoms with fixes.
Building a Retrieval-Augmented Generation Pipeline From Scratch
A walkthrough of the moving parts behind a production RAG system — chunking, embeddings, retrieval, and grounding an LLM's answers in real sources.
Fine-Tuning LLMs: LoRA vs Full Fine-Tuning
When parameter-efficient fine-tuning gets you 95% of the way there, and when you actually need to update every weight.
Prompt Engineering Patterns for Production GenAI Apps
The prompting patterns that actually hold up once real users, edge cases, and cost constraints show up.