What I keep learning about GenAI, RAG, agents and MCP, and notes from the things I build. I write when something finally clicks, or when something breaks in an interesting way.
Newest first. Filter by topic.
15 posts
DeepSeek open-sourced the inference engine behind its own API. What DeepInfer is, how prefill and decode are split, why KV caches live on SSDs for days, how DSpark speculative decoding works, and what the benchmarks show.
Running speech recognition, text-to-speech, voice cloning and a local language model on one 16 GB MacBook with Apple's MLX: the architecture, the real benchmark numbers, the quantization results, and the sharp edges.
Embeddings, cosine similarity, hybrid search and reranking, explained with runnable Python. Plus the failure cases that send good questions to the wrong page, and how to measure them.
Long context was supposed to kill retrieval. Then agentic search was. Why RAG keeps coming back, what IBM's 'Is RAG Still Needed?' gets right, and how to write docs retrieval can find.
How to build full-stack products with AI without losing control of the codebase: plan the product, map the system, and hand off to an AI IDE one tested slice at a time.
I kept pasting messy folder trees into AI chats, so I built a Rust CLI that prints a filtered, annotated, token-counted project map made for context windows. How it works, how it ships to npm, and what I learned.
A credible AI portfolio should prove the work is real: architecture, tradeoffs, links, metrics, and failure modes matter more than buzzwords.
Every developer has killed the wrong process to free a port. killit is a cross-platform CLI that checks what it's about to terminate before it does, with a three-level safety policy, graceful shutdown, and predictable exit codes.
What a year of building research assistants, alert-triage pipelines and coding workflows taught me about agents: when to use them, when not to, and the guardrails that matter more than the model.
Why I run language models on my own machine, the memory math that decides what fits, and how to pick a model size, all written while memory prices were climbing faster than my GPU fans.
How SynergyLearn is built: an AI learning platform with roadmaps, an in-browser IDE, sandboxed code execution and Gemini hints, where every external service is allowed to fail. Plus a rewrite of its setup guide.
Most DSA resources explain the answer. DSA Karle shows the algorithm running, step by step, from lesson content written as typed JSON. Here's how it works and why.
Building Email Digital Twin: a Chrome extension that learns how you write from your sent mail and drafts four replies in your own voice, with a demo video.
A practical guide to how generative AI apps actually work: what happens in one LLM call, how RAG grounds answers, what MCP plugs in, when you need an agent, and how to use each one in an ordinary workday.
A field guide to the ways language models go wrong (hallucination, stale context, drift, flattery) and the habits that keep AI-assisted writing accurate.