PinnedKavishka Kalhara·Nov 4Hybrid RAG: Why Your Retriever Needs Both Keywords and VectorsAn advanced, end-to-end guide to building retrieval that actually finds the right context: fusion math, rerankers, graph hybrids, code, and…A response icon1A response icon1
Kavishka Kalhara·11h ago10 GTA 6 Secret Locations Players Found in the First 48 HoursUFOs over Kelly County, a bunker with a code on a fire tower, and a house where someone figured out they live in a video game. Here is…
Kavishka Kalhara·1d agoChoosing a RAG Architecture Is the Hard Part, So I Built a Tool for Itragpick 1.0 is out: pick the right RAG architecture for your system, define it as code, and prove it works on your data.
Kavishka Kalhara·3d agoBuilding a Production-Grade BPE Tokenizer from Scratch in PythonByte-level BPE, GPT-4-style pre-tokenization, a training loop that runs 100× faster than the naive version, special-token safety…
Kavishka Kalhara·Nov 11Hearing the Human Through the NoiseHow noise cancellation works in AI voice assistants, and why it decides whether your voice agent feels smart or brokenA response icon1A response icon1
Kavishka Kalhara·Nov 4LLM Inference Optimization: The Complete Guide to Serving Models Faster and CheaperFrom roofline math and KV cache budgets to quantization, speculative decoding, disaggregated serving and a hands-on vLLM tuning workflow.
Kavishka Kalhara·Oct 24Your AI Agent Just Deleted a Failing Test. Here’s an Architecture That Would Have Stopped It.Vīthi: designing AI agents with a 2,000-year-old map of the mind
Kavishka Kalhara·Oct 17The Last Human-Driven Car Just Flew. Tesla Says AI Had Nothing to Do With It.Elon Musk needed a live crowd to prove his flying Roadster wasn’t AI-generated. That’s the most revealing part of the whole story.
Kavishka Kalhara·Sep 25What Is SAP? A Practical Guide to the World’s Biggest Business SoftwareFrom a five-person startup in 1972 to the backbone of global business, what SAP does, why it matters in 2026, and where it’s heading.
Kavishka Kalhara·Aug 23Stop Building RAG Pipelines for Documents That Fit in Context Cache-Augmented Generation is boring…You Built a Search Engine to Answer Questions About 400 Pages