AI Engineer · Voice AI · Infrastructure
Hi, I'm K. Boopathi. I build voice agents and LLM systems and keep them running once real users show up.
I started in full stack, got interested in DevOps along the way, and moved onto the infra and DevOps team. That turned out to be the useful half of AI engineering: the model is rarely what breaks. This site is where I keep what I learn. Posts from things that broke in production, notes from everything I explore on weekends, and a graph that connects them.
Case files
I started writing so the lessons from real work wouldn’t stay in my head. These are incidents I actually sat through, written up while the details were fresh. They read like detective stories because that is how they happened: a symptom, a trail of confident wrong guesses, and eventually the culprit. Usually me.
One innocent-looking command on a live collection of 200 million records. Connections spiked 2,900% and the whole database stopped negotiating.
MongoDB · Self-inflicted The AI that queried a million documentsOur agent called a tool with empty parameters. That became find({}) across a million documents inside a 1GB pod. Trusting the model was the real bug.
I scanned a cashback QR code out of curiosity. One query parameter later, I was looking at a quarter of a million people's personal data.
Security · Curiosity I blocked the attacker. It was me.A thousand requests a second hitting the landing page. I started banning IPs like a digital bouncer and banned our own internal network.
Incident · Humbling A million NXDomain errorsKubernetes was quietly generating over a million failed DNS lookups. The culprit was five characters of config nobody had ever read.
Kubernetes · DNS Heroku to Kubernetes, the honest versionEvery mystery the migration handed us, collected in one place: cold starts, race conditions, probes that lied, and jobs that restarted themselves.
Migration · Series hubWhat I go deep on
The work I’d point at first. Voice agents are where I’ve gone deepest everything else here is something that work pushed me into.
Twilio media streams, sub-500ms latency budgets, STT and TTS trade-offs, turn-taking, prompt caching and what it actually costs. My most complete original work.
Deepest expertise Doubling a cache hit rateA 7,200-token prompt, a 1,024-token cache floor, and first-turn hit rates that went from 24% to 62% on outbound calls.
Performance Inside ElevenLabs' agentsI pointed their conversation agent at my own LLM proxy to read the system prompts they send. Here's how it actually works.
Reverse engineering Prompt engineering, mechanicallyNot "try this prompt". How tokens, context windows and post-training actually shape what you get back.
From first principles The non-English LLM taxHindi costs roughly five times the tokens English does. At 10,000 messages a day that's an $82,000 gap nobody budgeted for.
LLM economics Four years at a startupIntern to team lead to infrastructure. The senior who told me "I am the documentation", and everything else nobody warns you about.
CareerNotes
I’m the guy who spends his weekends studying something new, and I explore enough of it that I used to forget most of it just as fast. So this is the place I built to remember every concept I’ve worked through, kept where I can find it again instead of learning it twice.
Most of it is organised by topic. Some of it is a rough thought I wrote down before it disappeared, and I’ve left those in as they are.
Browse all notes or start from a topic:
Second brain
AI is the field I explore most, and something worth knowing shows up in it almost daily new models, new tooling, new approaches. Keeping that in my head stopped working a while ago.
Plain notes weren’t enough either, because the useful part is usually how one idea connects to another. So this is built the way a brain stores things: every idea is a node, wired to the nodes around it. When something looks important, or just good to know, it goes into the graph rather than into a file on its own and when I come back to one node, its neighbours come back with it.
Two brains so far. Open the graph and follow the links:
The bigger of the two. Latency budgets, STT and TTS, turn-taking, telephony what I've learned building real-time agents, with the connections between it all.
AI engineeringNewer and still growing. LLM behaviour, evaluation, retrieval, agent patterns and the tooling around them.
Let’s connect
I’m an introvert right up until the code breaks then I’m completely in. If you’re building voice agents, chasing something strange in production, or you think I’m wrong about something here, I’d like to hear it.
The fastest way to reach me. Tell me what you're building and I'll reply.
GitHubThe code, the half-finished side projects, and commits with timestamps I'm not proud of.
YouTubeFor the things that are easier to show than to write down.
Dev.toThese posts, mirrored, if you'd rather argue in a comment section.
MediumSame writing, for people who live over there.
Buy me a coffeeOnly if something here saved you a bad night. Never expected.