Madhu · 18 essays in 4 series
Writing
Notes from building things: how retrieval is measured, how it is made better, and what it took to ship a few small apps. Each series reads in order.
Measurements5 essays
The four metrics that tell you whether a RAG or LLM system actually works — each one explained on its own. the whole series · map, cheat sheet, cards
1What is Context Recall?Of all the facts the right answer needs, how many did retrieval fetch? Recall is the ceiling everything else sits under.7 min2What is Context Precision?You fetched the right chunk — but is it near the top where the model reads, or buried under noise? That is precision.7 min3What is Faithfulness?Did the answer stick to what was retrieved, or quietly invent some of it? Faithfulness is your hallucination detector.7 min4What is Answer Relevancy?A response can be perfectly grounded in real context and still answer the wrong question. Relevancy catches the helpful-sounding miss.7 min5What is Retrieval Hit-Rate?Did you fetch any right doc at all?6 minRAG6 essays
Retrieval-Augmented Generation, built naive and improved one measured step at a time — Novice to Ninja. the whole series · map, cheat sheet, cards
1How RAG actually works, start to finishThe map for the whole series: a naive RAG system improved one idea at a time, every step measured on the same hard test. Start here.15 min2How do you actually measure a RAG system?The scorecard the whole series leans on — split RAG into retrieval and generation and score four meaning-aware metrics on a frozen golden set.12 min3How does chunking change RAG?The quietest knob in RAG and the cheapest improvement in the series. Change nothing but the size of the pieces, and beat the baseline on every metric.10 min4How query rewriting helps (and hurts)When ranking cannot help, change the question itself. Recall finally broke its plateau — with a latency bill worth being honest about.11 min5How does contextual retrieval work?Glue a one-line "you are here" note to each chunk before embedding. The biggest win in the series — and it let us delete a whole component.12 min6How does a semantic cache work?Answer once, serve instantly. Cut average latency 92% on repeated questions without touching accuracy — and one knife-edge setting to watch.11 minModels in the browser2 essays
Gemma 4 running inside a browser tab on WebGPU: no server, no API key, no credit wall — and what that does to the frontend job. the whole series · map, cheat sheet, cards
1There's an LLM Running in Your Browser Right Now — No Internet, No API, No CreditsEvery app shows you that 0/3 credits screen. Gemma 4's TurboQuant + PLE make it possible to run an LLM entirely in the browser — no internet after…23 min2UI Out of Thin Air — What Frontend Engineers Will Actually Do After Gemma 4A lot of frontend engineers are quietly worried that Claude, v0, and Cursor are eating the job. They're worried about the wrong thing. Gemma 4 just…16 minBuild notes5 essays
Short write-ups of things I built, most of them in a weekend. the whole series · map, cheat sheet, cards
1Thinking in Agents: The New Mental Model Every Developer NeedsWe used to write code that told computers what to do. Now we're writing code that tells agents what to think about.8 min2OptionMantra: VIX Data Pipeline - A Practical, Everyday BuildBuilding a cost-effective VIX data pipeline with Python, SQLite, and Google Cloud Storage. From $500/month to $2/month.9 min3Introducing WordStash: Your Personal Vocabulary CompanionBuilding a dead-simple iOS app for capturing and learning new words. From copy-paste to AI-powered definitions in seconds.8 min4Spec-Driven Development: Building ContentFlow CMS in 2 Days with AIHow I built an entire CMS—SDK, portal, server, deployment—in 32 hours by treating specifications as the primary artifact and letting AI handle 90% of…18 min5Building ContentFlow CMS: A Journey into AI-Powered DevelopmentHow I built an entire content management system using Level 4 AI development—SDK, CMS, server, and deployment—in just 32 hours.14 min