Ruchit Suthar
AI writes the code. I still make the call.
Fifteen years in, the question I ask about my own work has changed three times. First it was did it ship. Then is it right. Now it's should we build it at all. I'm a software architect — I own the tech stack, the infrastructure, the ADRs and the trade-offs on products where a wrong call costs real money and real users. This site is that third question, written down.
What the work actually is
Six things I get called in for. Everything written on this site comes out of one of them.
- System design under real constraints
- Architecture for products where a wrong call costs real money and real users. Tech-stack and infrastructure decisions, the trade-offs nobody else wants to put their name on, and ADRs written to still be readable after the people who wrote them have moved on.
- AI in production, not in demos
- Prompt engineering → context engineering → MCP → RAG → agentic systems, shipped inside products that have users and an on-call rota. The interesting question was never whether it works in a notebook.
- Engineering leverage
- A custom agent-orchestrator that walks a build through each SDLC step with the review gates intact — MVPs land about 3× faster. Leverage is a system you build, not work you do faster.
- Teams that hold
- 25+ engineers led and more than 1,000 technical interviews conducted. Who you hire, how you level them and what you let them own is the most leveraged architecture in the building, and the only part with no rollback.
- Quality as a system
- Review discipline, definition of done, and the gates that turn output into something still standing two reorgs later. Quality is not a phase at the end; it is a property you design in or pay for twice.
- Scale that has already broken
- 1,500 requests a minute at peak, high-traffic data platforms, distributed teams across five time zones. Every pattern here is one I have watched fail before I recommended it.
Output, then quality, then judgment
Three phases, each one added to the last rather than replacing it. The parts that cost me are included, because those are the ones that actually changed how I work.
- 01Output
Ship it, keep it running, own the outcome.
I learned systems from the metal up — server and network config, log analysis, incidents on a platform taking 1,500 requests a minute at peak. Before I was any good at writing code I was good at keeping it running. That is why architecture came early: you already know what breaks at scale when you have been the one woken up to fix it.
What it cost
I hired for a story instead of the signal in front of me. The team carried that gap, not me. Sympathy is not a hiring signal, and I learned it the expensive way.
- 02Quality
Output that survives contact with time.
Then the bar moved. Real process at every stage, code reviewed before it reached production, clean code, design patterns, ADRs. Work I used to ship in a week took a month, and I resented it until I understood what the month was buying. Output got me hired. Quality is what made me an architect.
What it cost
A product that missed the quality bar cost three months of rework. The corner you cut to hit a date is a loan, and the interest is paid by whoever is holding it later.
- 03JudgmentNow
The decision behind the code — and now, behind the machine.
What is worth building at all, what a trade-off actually costs, and what AI should and should not be trusted with. AI did not replace judgment; it moved it. The scarce skill is no longer producing the code — it is deciding what deserves to exist and checking what came back. The bottleneck is not generation. It is review.
What it cost
A function I left looping quietly burned $5,350 in a month, unwatched. And I once stayed kind with someone who needed the truth, until they ran out of runway. Both were judgment, not knowledge.
Six positions, and where each one is defended
Principles on an about page are free. Each of these is anchored to something that went wrong and to the article that makes the full case — disagree with one and there is somewhere to take it.
The right architecture is the one your on-call can fix at 3am without you.
Elegance gets graded in the design review. Recoverability gets graded at 3am, by someone tired and alone. Every clean-on-paper decision I regret passed the first test and failed the second.
Read the argumentOver-engineering is a confession that you didn't understand the problem.
I built multi-tenancy and elaborate abstractions for millions of users. It shipped to three. Elaborate architecture is what you reach for when you cannot yet name the real constraint.
Read the argumentAI didn't make judgment obsolete. It made judgment the only scarce skill.
Neither the doom take nor the ten-x take. Generation got cheap and review did not, so the value moved to the one step that never automated.
Read the argumentBeing nice is deferred cowardice. Kindness is telling people the truth early.
I softened hard feedback to protect someone's feelings in the room. The interest on a withheld truth is always paid by the person who never got the chance to act on it.
Read the argumentA system's real cost is the code you forget is running.
$5,350 in a month, from a function nobody was looking at. Cost is not a finance problem that arrives later — it is a property the architect designs in or leaves out.
Read the argumentLeverage is a system you build, not work you do faster.
The workflow that unblocks a whole role beats anything I could have personally shipped that quarter. Optimising your own throughput is the last place to look for a multiplier.
Read the argumentFive reasons people land here
128 articles, 10 pathways, 6 reference documents and 3 calculators — a lot to walk into cold. Find the sentence that sounds like your week and start there.
You're senior, and the ladder stopped.
The IC track flattens at senior and the management door opens at the same time. Nobody tells you the trade you are actually making.
You started leading last quarter.
Promoted for being good at the work, into a job that is mostly people, politics and cadence. The job description lied by omission.
Your system outgrew its design.
You already know the failure modes. What you need is the sequenced path out, and numbers current enough to put in a design doc.
You're putting AI into work that matters.
Past the demo, into a product with users and an on-call rota. You want the failure cases named before the benefits.
You're signing off on decisions you can't evaluate.
A founder without a technical cofounder, buying the build or managing engineers you could not interview properly. You need a disinterested read, not a vendor's.
Questions people actually ask
Not what this page is. What a reader weighs before quoting a personal site in a design doc, a deck, or a decision.
- Who is Ruchit Suthar?
- Ruchit Suthar is a software architect and technical leader with over 15 years in production software. He works on system design under real constraints, engineering leadership, and AI in production engineering; he has led 25+ engineers, conducted more than 1,000 technical interviews, and writes ruchitsuthar.com.
- Is any of this AI-generated?
- AI tools are used for editing, formatting and SEO. Every idea, opinion and technical recommendation is authored by Ruchit Suthar from work he actually did. The full editorial policy, including how corrections are issued, is published on the site.Editorial policy →
- Can I quote or cite these articles?
- Yes — quote freely with a link back to the original. Every article also publishes a plain-text llms.txt alongside it, so an AI assistant can read the source rather than a summary of a summary.Site llms.txt →
- How often is new work published, and how do I follow it?
- New articles go out by email when there is something worth sending, with no digest padding. There is a full RSS feed for anyone who would rather not hand over an address.Newsletter →RSS feed →
- Does Ruchit take mentorship sessions?
- Yes, and they are free. A limited number of one-to-one sessions for engineers moving from writing code to owning decisions, first-time leads, and people weighing the architecture track.Mentorship →
128 articles · 10 pathways · 6 reference docs
If any of that was useful, the rest of it is free too.
New work goes out by email when there is new work — one piece, no digest padding, no drip sequence. Or take the RSS feed and never give me an address.
Output got me hired. Quality made me an architect. Judgment is the only part the machine has not taken, and it is the part nobody can hand you — you pay for it in outages, rework and one conversation you had too late.