https://www.realfast.ai/sumanth-why-realfast

Sumanth · The Janitor Recruitment, branding & marketing, people ops, India office operations, basically anything that comes my way. ~8 min read

This is a long read, and that's on purpose. It's where I lay out what realfast actually is and what you'd be signing up for, the honest version, not the careers-page one. If you're weighing this as your next job, read it end to end before you apply. And if sitting through a long read already feels like too much, this probably isn't the right place for you, better to find that out now than three weeks in.

Landed here before applying? You can apply directly at join.realfast.ai. Connect your agent and we'll take it from there. Please complete your application or agent setup before expressing interest below.

The people behind realfast have done this before, at a scale you don't see in India often.

Sidu Ponnappa and Aakash Dharmadhikari, the founders of realfast, both have more than two decades of experience in tech and services. They ran a consulting company that got acquired by Gojek (now GoTo), Indonesia's conglomerate company, with 250M+ users.

Aakash was Group CPO. Sidu was MD of India and CIO of Singapore.

To put that in numbers, in 2016, post-acquisition, the team was around 30 people. By 2021 it was 800+ engineers in India and almost 3,000 across Southeast Asia. I worked closely with both of them between 2016 and 2021, so this isn't second-hand. I've seen how they build, how they assess, how they rebuild. More importantly, I've seen what breaks when you try to scale enterprise software, and they've seen it ten times worse.

That's the founder context. Now here's why they started realfast.

They started in 2022. ChatGPT launched the next month. They rethought the whole plan.

The original plan was a high-end firm, built the way Sidu and Aakash always wanted to build one: real strategy, deep engineering, no body-shopping. Then GPT happened, and within a few weeks it was clear the thing they'd meant to sell, smart people billed out by the hour, was on its way out. Not "disrupted." Done.

If you've been watching the market, when Opus 4.6 dropped a couple of months ago, the services stocks tumbled. Public markets finally caught up to what we've been saying since 2023. 2022 already feels like a different era.

Every part of that sentence means something specific. Here's what each bit actually is.

Applied-AI. We're not a research lab. We take what's shipping now, the current frontier, and put it to work on problems that have a real cost attached. Most of the work is closing the gap between "this works in a demo" and "this runs in your production system." That gap is bigger than people expect, and it's most of the job.

Strategy and engineering. Most firms do one. Strategy houses write the deck and hand you a plan you can't run. Dev shops build exactly what you spec, without asking whether the spec is right. We try to do both with the same people, the ones deciding what's worth doing are the ones building it. It's harder to staff and harder to sell, and we think it's the only version that works now.

Software-driven enterprises. We work with companies where software is the business, not a support function. That's where this changes the P&L instead of the org chart.

Urgent, high-stakes problems. We're not here to groom your backlog. We take the expensive, high-stakes problem you've been putting off. The flip side: if the problem isn't urgent, we're overkill and a bad fit, that's fine, we'll say so.

Products and platforms that move the metric. We don't hand over a deck and an invoice. We leave something that runs, that you own, tied to a business number. And we hold ourselves to that number, if it doesn't move, we didn't do the job.

That's who we are. The rest of this is how we actually do it, and where it gets hard.

Two things make this place different. Both are uncomfortable for most people.

One, judgement over code speed. Our engineers spend ~70% of their time on architecture and judgement. ~30% writing code. That's not an accident, it's the whole point. We hire people who can direct agents, not just out-type them. AI flipped the ratio. The value now is in knowing what to build, why, and when to throw it away. Not in typing it out.

Two, small team, absurd output. Four or five exceptional engineers with AI leverage out-shipping teams that are 3 to 5x their size. That's not aspiration, that's how the current projects are run. If you're used to being the smartest person in a 40-person team, this will feel weird. The team is tiny on purpose.

He says we don't want majdoors (wage labour). We want karigars (craftspeople). Craft matters. Taste matters. Integrity matters. If that lands for you, keep reading.

A customer spends a dollar with us, they see two back, in days, not quarters.

We don't sell scope. We don't sell hours. We sell movement on a number that matters to the business.

The old way, the way services firms have always worked, is to lose money on discovery and make it back on delivery. Senior people spend weeks, sometimes months, in client meetings, reviewing docs, reverse-engineering systems, capturing requirements. All of it is cost. Slow, expensive, and the knowledge ends up trapped in someone's head.

The industry's "AI response" was to give engineers Copilot and hope. That speeds up delivery, sure. But it leaves the expensive part, discovery, completely untouched. The real bottleneck isn't code speed. It's time-to-monetisable-code: the total time and cost from contract signing to the first deliverable a client is actually paying for. Everything before that is cost. Everything after is revenue.

That's the metric we built the company around. The pitch we make to enterprises is, give us 2 to 4 weeks. We'll carve a problem out crisp, attach an operating metric, and ship a product or platform against it. If we don't move the number, you're not happy with us, don't pay us bro.

That line is Sidu's, verbatim. It's also actually true. It's why customers keep signing on.

There's no such thing as an "agentic workflow." That's just bakchodi.

Look, go on LinkedIn right now and search "AI transformation". It's a swamp. Everybody is selling. Almost nobody is delivering. Most people are using AI to do mediocre work faster and calling it transformation.

Here's what we actually believe: if you can't deliver superhuman speed AND superhuman quality, it's not transformation. It's marketing.

Sidu's test, "I can take an agent, point it at a 5 million line codebase, and find things. Humanly possible? No. Agentically possible? Yes. If I can do it, and I'm just a developer who codes for a living, you should also be able to do it. If you can't, you're a scammer."

That's the bar. If you've been hiding behind "AI" titles without actually shipping with agents, this place will eat you alive in the first month. If you've been quietly out-shipping your team because you actually know how to drive a model, you'll fit right in.

Real numbers. No fluff.

Sidu, in his own words, on the timeline. Easier to hear it from him than from me.

If you want a fully de-risked Series-D rocket ship, we're not it. If you want a place where the founders have done it before, the money's in the bank, and the customers are real, yeah, we're it.

AI doesn't make work easier. It makes it harder, and massively more productive.

2x the effort, 10x the output. That's the trade. We're early-stage. The thesis updates every three months because the model capabilities update every three months. Things change fast. It's not always clean.

This is genuinely not for people who want stability or want to do tomorrow what they did yesterday. If that's where you are in your life, totally valid, take a job at a stable big tech, do good work, go home.

But if you've been watching the market and thinking "the way things work right now is already obsolete, I want to be on the side that's building what comes next", yeah, we should talk. That's the whole reason this company exists.

Attention span to consume large amounts of text is table stakes here. With agents, it's the actual job.

We don't read your code first. We read how you worked with the AI.

Here's a thing most companies haven't caught up to yet. The take-home used to be the test. Then models started one-shotting it, cleanly, and honestly better than most of us would have. The whole thing rested on one assumption: good output means a good engineer. That assumption is gone. Output is table stakes now.

So we spent the last six months running experiments on that single step, the take-home in the age of coding agents. And we found something that's changed how we hire.

We now ask candidates to submit their JSONLs: the full transcript of how you worked with the AI. A recruiting agent makes sense of the ~25MB of JSON each one generates, and we read that more closely than the code.

It's the old math exam. You don't get the marks for the right answer. You get them for showing the right method.

So what's the right method? We've turned it into a rubric we now hire against. Three things, in plain English:

»AI Collaboration. Do you get more out of AI than the next person?

»Understanding. When you hit something you don't know, are you learning, or just collecting answers?

»Rigor. Can people trust what you ship?

This has flipped our decisions. We've accepted people whose output was only good enough, but who worked with AI exceptionally well. And we've rejected people with stellar submissions who turned out to be passengers, while the AI did the real work.

Knowing what to look for was the easy part. Mapping every candidate across this rubric from their raw JSONLs has been the single most challenging AI evaluation project we've taken on.

And it reaches well past hiring, into a question every leader is about to face: when a model can one-shot the output, how do you run a performance review? Are you evaluating the human or the agent? For as long as we've graded knowledge work, we've been stuck on the artifact, the doc, the PR, the deliverable that comes out the end. Process was invisible, so we reached for proxies: hours logged, tickets closed, a peer's gut sense of whether someone was "sharp." The agent logs change that. Read them and you get a window into the person's mind: where they pushed back, what they anticipated, whether they steered toward the hard parts or just accepted the first plausible answer. You can see whether someone has real judgment or is a passenger. That's a depth of signal we've never had before, and it's already changing how we review, mentor, and decide who can take on a whole new class of harder problems.

Anyway, that's my pitch. If any of this resonated, the next step is the form on the interview kit. It's two minutes. We've moved past recruiter intro calls; everything you need to decide is on this page and the kit.

If something didn't land, or you've got a sharp question, just reply. I read every one. I'm the Janitor, remember.

Oh, and one more thing. This one's me, in case you were wondering what kind of person you'd be talking to.

The Janitor · realfast

P.S. If you're in, fill the form now while it's fresh. Even a quick yes lets us start moving, we'll send the assignment shortly after, and you can take it from there.

P.P.S. And if you're still on the fence, keep scrolling. Trust me, there's more below where I explain the rest, what an FDE actually is, the FAQ, the reading list. More than I'd cover on a call.

Express Interest Now

THE PLATFORM · EXOCORTEX (AKA EXO)

Plan, code, review, and verify agents working in concert against real customer codebases.

Point an agent at an enterprise monorepo. Find what humans can't. Sidu's bar. If you can't, you're a scammer.

Every project ships against a measurable business metric. The platform tracks deltas, not story points.

Prompt libraries that learn from production traces. Captured engineering judgement, replayable.

THE TEAM

THE ROLE · EXPLAINED

You operate in cockpit mode.

Lead client engagements end-to-end. Discovery, prototype, production, same pair of hands. Customer in the room, agents on the keyboard, you steering. Days, not months.

Lead from the code Own the outcome Customer-facing

Real fluency in agentic systems.

This isn't "have you tried ChatGPT." We expect you to have shipped something real with agents: orchestration, tool use, retries, evals, the messy stuff. If you can't drive Claude Code or Codex through a non-trivial change today, this role will hurt.

Claude Code / Codex Agent orchestration Evals & guardrails LLM tool use

"AI is the senior engineer. Humans are the architects."

THE INTERVIEW PROCESS

One-minute form. Just the essentials: name, email, phone, the role you're keen on, your notice period, and your expected comp so we can align on budget early.

Fill the Form

AFTER THE FORM · MANDATORY

Mandatory: complete this screening after the form. Your assignment is sent once both are done.

READING MATERIAL

Quick note before the list. Reading volume isn't a hobby here, it's the job. Agents get you exactly as far as your input layer: what you've read, what frames you've collected, what mental models you can reach for when the spec gets ambiguous. The people who plateau fastest are the ones who stop at technical docs. The ones who keep getting sharper read widely, and most of them read sci-fi. Not because we require it, but because the authors who write it (often physicists, mathematicians) end up handing you the metaphors you'll need a decade before the rest of the industry gets there. If reading 30+ pages a day in your own time sounds heavy, this probably isn't the right place.

{{ rd.label }}

{{ rd.iconEl }} {{ rd.title }} {{ rd.author }} {{ rd.kind }}