Writing

Field notes from the work.

Essays on cloud architecture, AI engineering, and what training thousands of people has taught me about how we actually learn hard things.

October 29, 2026 · 14 min read

Why an AI Sent 155 People Someone Else's Data: How AI Memory Gets Reused, and How to Check Yours

In August 2026 an AI model on Cloudflare Workers AI returned 155 answers that contained other customers' data. The cause was one number that didn't match: the model accepted 1,048,576-token prompts, but its GPU memory held 681,280. Here's how AI memory gets reused between requests, why that leaked, what I found when I tested how far caches are shared on my own Azure account, and the checks to run on yours.

#ai-engineering#security#llm-inference#azure#incident#experiment

October 28, 2026 · 14 min read

Is AI's Production-Ready Code Ready for Production? I Tested 100 Services

I asked five AI models for a production-ready link-preview service, 20 times each, then built every answer in Docker, attacked it and scanned it with 15 checks written before the first result. 17 of 100 passed all 15. 23 didn't work at all, and 28 of the 77 that worked fetched an internal address and returned its secret. Here's what failed, why, and the checks to run before AI-written code ships.

#devops#ai-engineering#security#docker#kubernetes#experiment

October 25, 2026 · 18 min read

Hindi In, Urdu Out: I Tested Voice AI on Indian Languages

I sent the same recorded sentences in Hindi, Bengali, Marathi, Tamil and English to three speech-to-text services and scored every transcript. With the language set to Hindi, one service returned 50 of 390 Hindi clips in another script, and 1 in 3 when the clip was under 8 seconds. Here's what each service got right, what one line of instruction fixed, and what an hour of audio costs at list prices.

#ai#speech-to-text#voice-ai#hindi#indian-languages#experiment

October 25, 2026 · 14 min read

Why Hindi Pays More: Inside the Tokenizer

Part 2 of the Hindi token tax experiment. I followed one sentence from the keyboard to the model, counted every stamp in 10 tokenizers' vocabularies, and trained five small tokenizers that differ only in how much Hindi they saw. The cause is how much of a script the tokenizer was built from, with bytes setting the starting line.

#ai#llm#tokens#tokenizers#hindi#indian-languages#experiment

October 22, 2026 · 15 min read

Hindi Pays a Tax on AI: I Measured It Across 10 Tokenizers and 11 Indian Languages

The same text costs more tokens in Hindi than in English, so it costs more money, fills the context window faster and takes longer to answer. I measured how much on 1,012 professionally translated sentences across 10 real tokenizers, found that the tax has fallen sharply, that other Indian languages pay more, and that typing Hindi in English letters now costs more on modern models, not less.

#ai#llm#tokens#hindi#indian-languages#cost#experiment

October 20, 2026 · 26 min read

The VM Is One Line on the Bill: Learn Real Infrastructure on a $12 Droplet

A field guide for learners: why a running app costs more than its VM, how to list every resource before you price anything, how to keep a learning project from becoming a surprise bill, and a hands-on DigitalOcean lab covering SSH keys, hardening, firewalls, a private network and a database the internet can't reach. With a fair, sourced comparison against AWS and Azure, and a safe way to let an AI assistant help.

#cloud#digitalocean#linux#ssh#networking#cost#beginners

October 19, 2026 · 24 min read

I Crashed My AI Worker Twice: Azure Service Bus for AI Apps

A plain-English guide to putting Azure Service Bus between your API and your AI model: load leveling, competing consumers, topics and filters, message design, claim-check, peek-lock and the dead-letter queue. Every claim tested on the free local emulator, including a real worker crash in both receive modes.

#azure#service-bus#messaging#ai-engineering#distributed-systems#experiment

October 9, 2026 · 12 min read

Big Embeddings in pgvector: The 2,000-Dimension Limit, Two Fixes, and What They Cost

Many embedding models now output more than 2,000 numbers per text, and pgvector can store them but not index them. A plain-English experiment report: the error, why it exists, half-precision indexes, shorter embeddings, and the cast that silently turns your index off. Measured with real embeddings and real questions.

#pgvector#postgresql#embeddings#rag#ai-engineering#experiment

October 9, 2026 · 8 min read

Is Your Vector Search Slow? 5 Fixes, Measured

Five common reasons vector search in Postgres feels slow, each tested on 5,000 real passages and 300 real questions: no index, fetching the vectors back, a new connection per question, the ef_search knob, and big vectors. Plain English, Indian memes, real numbers.

#pgvector#postgresql#vector-search#performance#rag#experiment

October 7, 2026 · 19 min read

Caching a RAG App with Redis: What to Cache, When It Goes Stale, and What Breaks

A plain-English guide to putting Redis in front of a RAG app. Learn cache-aside and TTLs, cache the three slow steps of a RAG request, watch a cached answer go stale when a help page changes, and see what happens when Redis is down or full. Every number is measured.

#redis#caching#rag#ai-engineering#azure#performance

October 7, 2026 · 18 min read

Redis Pub/Sub vs Streams for AI Pipelines: Broadcast the News, Queue the Work

A plain-English guide to the two ways Redis moves messages between services. When a help page changes in a RAG app, the re-embedding work goes through a Stream and the cache invalidation goes out on pub/sub. Real crashes, lost messages, duplicate work and timeouts, with every output shown.

#redis#streams#pub-sub#rag#ai-engineering#azure

October 5, 2026 · 14 min read

Debugging RAG: Check Retrieval Before You Touch the Model

A plain-English guide to why RAG answers go wrong. Split the system into a clerk (retrieval) and a writer (the model), run a five-minute test to find out which one failed, then try it on a small help centre with pgvector. Every number is measured.

#rag#ai-engineering#retrieval#pgvector#embeddings#debugging

October 4, 2026 · 21 min read

From Rows to Meaning: How Vector Databases Work, and How to Build or Add One

A plain-English walk from an ordinary database to vector search: indexes, embeddings, nearest neighbours, IVF and HNSW, one mental model, a vector database built from scratch in Python, and a step-by-step upgrade of Postgres with pgvector.

#vector-databases#ai-engineering#postgresql#pgvector#embeddings#system-design

October 2, 2026 · 56 min read

How a Kubernetes Cluster Actually Works: From kubectl apply to Packets on the Wire

Control loops, etcd, watches, the scheduler, the kubelet, Services and kube-proxy, traced end to end through one Deployment, with failure timelines for what happens when a node, the control plane or etcd breaks.

#kubernetes#architecture#system-design#controllers#networking#etcd

September 30, 2026 · 100 min read

Containers: From Source Code to Image Layers, SHA-256 Digests, OCI Artifacts, and Running Processes

What actually happens between docker build and a running container: BuildKit, OCI manifests, content-addressable blobs, registries, snapshotters, OverlayFS, namespaces, cgroups and runc, traced through one small application — with live, interactive diagrams for the parts you normally can't see.

#containers#docker#oci#linux#kubernetes#containerd

August 19, 2026 · 4 min read

The Scale Tax: what auto-scaling doesn't tell you about your next invoice

Scaling an app is the fun problem. Paying for it after the traffic leaves is the one nobody schedules a meeting for. A field record of how we built a code-execution engine that scales to 500+ concurrent users — and the auto-scaling traps that nearly kept it billing us long after the users went home.

#cloud#cost-optimization#aws#case-study

August 8, 2026 · 3 min read

Buying a domain: registrars, ICANN, and WHOIS

What you're actually paying for when you buy a domain, how ICANN accredits the registrars that sell to you, what ICANN Lookup shows about any domain — and why owning a name still doesn't point it anywhere. Part 2 of 5.

#course#fundamentals#domains

August 8, 2026 · 2 min read

How this blog is built

A quick tour of the stack behind aryantripathi.dev, and why it's deliberately boring.

#engineering#meta

August 8, 2026 · 3 min read

From localhost to a live URL, no domain required

What localhost:3000 actually means, why nobody else can open it, and how a free host gives you a real public URL the moment you connect a repo — Part 1 of a 5-part course on how a website actually gets from your laptop to the internet.

#course#fundamentals#hosting

August 8, 2026 · 1 min read

Welcome to the blog

Why this exists, what it's for, and what to expect from the field notes I'll be posting here.

#meta