devendra variya

gujarat board, hsc. then two years on a hardware and networking course at the institute of advance network technology. cables, switches, routers, windows server.

the course was not much help. i learned devops off youtube instead.

the useful part of those two years was accidental. because i learned networks first, the cloud was not a new set of ideas later. same ideas, running on machines i cannot touch, with a bill attached.

wanna know why that order mattered? here's the post:

my first paid job was $25.

overlapping with that i took a year as a junior devops at a small agency, on call for their client servers. that year is where i actually learned it. being responsible for something at 3am teaches faster than any playlist.

the work was small at the start. fix the deploy. move it off the one server it has been sitting on since 2016. replace the certificate that expired last tuesday.

p.s. an alarming share of professional devops is renewing certificates.

to get that work i moved my whole life onto someone else's timezone. i have been on a us clock ever since. it paid for everything on this page and it costs more than i usually admit.

most of the work came through upwork. 105+ projects finished, 100% job success score, mostly us and uk clients.

that second number now runs my business. it is the reason i turn down work i could probably do, and the reason i ask a client for their budget before i give them a rate.

i do not name clients in public. most of what i touch is somebody's internal system and it is not mine to advertise. the stories here are about the bugs, not about who was paying for them.

me and rahul, doing infrastructure work under one name instead of two. no office, no plan longer than the next project.

somewhere in there the certificates happened. hashicorp terraform associate 003. oracle cloud infrastructure foundations.

the thing that actually changed my work was not a certificate. it was writing infrastructure as code and realising the job is to describe a system, not to build one by hand and then spend years defending it.

the other half of that work is less glamorous. a lot of what i get hired for is making a cloud bill smaller, and almost none of it is clever architecture.

bca at ignou, while running the company. distance learning, which mostly means exams land in the worst week of whatever project is live.

i dropped it again. i never got interested in studying, and by then the skill was worth more to me than the paper would be. it is not finished and i am not going to dress that up.

p.s. i had been working in this field for four years before i enrolled. the order is backwards and i am fine with it.

we worked together for two years and most of it was good. i do not think that part was fake. it was the ending that changed what the whole thing means.

it cost me more than $50,000. i am not naming anyone and i am not telling it with a villain in it, because the part that was mine to own is that i did not check anything. i trusted, and then i kept trusting after i had reasons not to.

i do not count it as a loss. god was watching, and he has already given back more than it took. so i count it as the price of one lesson: do not overtrust anyone.

p.s. i still work on trust. i just verify now.

mypage.cc went out on product hunt. link in bio, except designed pages instead of a stack of grey buttons. astro, cloudflare workers, d1, r2.

then i audited my own product and it was ugly. pricing copy had drifted across five hand-maintained places. demo personas were indexed with self-referencing canonicals. the refund window was shorter than the free trial.

p.s. auditing your own launch two weeks later is free, and it stings.

wanna know more about that launch? here's the post:

screenfine fines you when you cross your own screen time limit. you earn the fine back with camera-verified pushups, squats or situps. $1 a week. expo, react native, native swift modules, convex.

apple approved the familycontrols entitlement, scoped to me managing my own device. that approval is the entire product. without it there is no app.

it won daily winner on daniellaunches. the loudest comment asked for scheduled lock windows. so now i am building scheduled lock windows.

here's what building it actually looked like:

envguard is the least glamorous thing i have published and the one i use every day. 8+ npm packages for environment variable validation, with a cli, a github action and a vs code extension.

why eight packages and not one:

shipping two of my own things in one year taught me something i did not want to learn: i make the same mistakes i get paid to find in other people's work.

p.s. i went looking for an algorithm update and found a file i wrote myself.

screenfine is pre-revenue. i put that in writing because a daily winner badge looks like money to people who have never launched anything, and it is not.

the spike ends in about a week. what stays is a link that ranks, one comment worth more than the rest of the day, and a product that now needs distribution work nobody enjoys.

i have never had runway. no savings to burn, no quitting to build full time. the client work pays for the products, which means the products get the tired hours and never have to earn rent.

p.s. pricing an app at $1 a week only makes sense if you are not desperate.

and then there is the half of shipping that nobody posts about. vat, consent banners, a refund window that contradicted my own free trial, a privacy policy describing a product that had moved on without it.

stacklance infotech llp, llpin acz-9112. registering a company is not a milestone, it is paperwork. the work that made it worth filing had already happened.

six years of freelancing, compressed into paperwork:

people stopped asking me to move servers and started asking for retrieval systems and agents. so most of what i build now is rag pipelines, orchestrators, and bots that answer questions out of a company's own documents.

the failure mode nobody warns you about: a rag system does not crash. it quietly says 'i don't have that information' to a question it can answer perfectly well, and everyone assumes the documents are bad.

here's the post about the two bugs that taught me that:

p.s. i'm a developer, not a writer. this is the fifth draft.