Why My Portfolio Lives in the Cosmos
There's a moment, right before you open someone's portfolio, where you already have a mental model of what you're about to see. A clean white canvas. A hero section with a name in large type. A grid of projects below it. Maybe a dark mode toggle in the top right corner. You've seen it a hundred times - not because designers lack imagination, but because the genre has converged on a template so thoroughly that the template has become invisible.
I didn't want that.
Not out of contrarianism, but out of a genuine belief about how attention works, and what it means to ask a stranger to spend their time understanding who you are.
Before I was an AI systems engineer, I was a designer.
Not in the casual sense of someone who picked up Figma and ran with it. I mean the kind of designer that gets shaped by deadlines and editorial standards and the specific humiliation of a layout that almost works but doesn't quite. In school, I was lead designer for both the class magazine and the school magazine for four years - the class magazine won Best Magazine two years running. In Class 11, I was designing and editing for the Times of India student edition. Through college, I ran design for the BMSCE IEEE student branch and did graphic design for Outlawed for over two years each.
That's nearly a decade of thinking about composition, hierarchy, negative space, typographic rhythm, and the invisible logic that makes a reader want to turn the page before they've consciously decided to. That eye was formed long before I wrote my first neural network or debugged my first RAG pipeline.
I say this not to credential-stack, but because it's the honest explanation for why this portfolio looks the way it does. The design choices here aren't an engineer reaching awkwardly into unfamiliar territory. They're the result of someone who has spent a long time thinking about what makes visual communication actually work - and who happens to also build AI systems for a living.
The cosmos-themed, animation-heavy, almost aggressively immersive portfolio is what you get when those two things meet.
Here's something I've thought about a lot, both as a designer and as someone who works on AI systems: most decisions are made before the deliberate part of the brain gets involved. A user who lands on a portfolio page makes a subconscious judgment in the first second - is this worth my time? - and the answer to that question isn't determined by the quality of the projects listed or the eloquence of the bio. It's determined entirely by what they feel when they arrive.
This sounds like a marketing truism, but it has real implications for how I thought about building this site. If I wanted someone - a recruiter, a collaborator, a curious engineer, anyone - to actually read about what I do and what I care about, I first had to give them a reason to stay. Not a clever reason. Not a rational reason. A visceral one.
Every page I designed for a magazine had to earn its reader before they committed to the article. A feature spread that didn't arrest the eye in the first glance got skipped - regardless of how good the writing was. Portfolios work exactly the same way, and almost none of them seem to know it.
So I made the choice that most engineers wouldn't: I went design-heavy. Deliberately, unapologetically, almost excessively so. Layered backgrounds. Depth of field. Flowing grain. Smooth scroll behaviors that took more debugging time than some features I've shipped at work. A comet animation threading through the experience timeline that I rebuilt three times because the timing felt wrong.
Was this over-engineered for a portfolio? By the metrics of conventional wisdom, absolutely. By the metric of did it make someone pause and look closer - I think it works. The design is the first argument I make about myself before a single word of copy is read.
The problem with most portfolios - and I say this with genuine respect for the people who built them - is that they don't feel alive. They feel like well-organized documents. And a well-organized document is a thing you consult, not a place you inhabit.
Years of editorial work taught me the difference. A magazine spread isn't just content on a page - it has atmosphere. There's a reason why the best magazine pages feel like environments: the interplay of type and image and white space creates a kind of pressure, a sense that the page has intention and that the reader is inside something rather than just consuming it.
Every design choice I made on this site was in service of that same goal.
The multi-layered parallax background isn't decoration. It's spatial depth - the kind that signals to your peripheral vision that there is more world beyond the edge of the screen. The flowing grain in the background mimics the subtle texture of a real environment, the way a room feels different from a void. The comet animation on the experience timeline - a small glowing trail arcing across milestones - turns what is otherwise a static list of dates and company names into something that feels like a journey being traced across time.
The cursor is the smallest of these details and maybe the most precise. The native OS pointer is gone, replaced entirely with the International Space Station - its truss, its solar panel arrays, its communication dish rendered in line by line. When hovering over something interactive, the dish emits arcs, five of them staggered so two are always mid-travel. This is technically a hover state. But the ISS is a system: a complex, multi-component structure doing one difficult thing extremely well under conditions that would destroy most designs. It belongs in this environment. And when the dish fires, what is nominally a UI affordance becomes something closer to a micro-narrative. The signal is going out.
These aren't tricks. They're the difference between a room with furniture in it and a room that feels like someone actually lives there.
I chose the cosmos specifically because space does something no other visual metaphor quite achieves: it makes you feel small and curious at the same time. Looking at the night sky, you don't feel anxious about how much you don't know - you feel drawn to it. That orientation, that posture of curious engagement, is exactly what I want a visitor to bring to reading about my work. Come in curious. Stay because something keeps catching your eye.
But the cosmos isn't just an aesthetic choice. It's the most honest metaphor I've found for the kind of work I actually do.
When something goes wrong in an AI system - and something always goes wrong in an AI system - the failure is almost never a single clean bug with a stack trace that points directly to the problem. It's a confluence. It's a retrieval component that returns slightly off-context chunks, combined with a prompt that was written for a subtly different input distribution, combined with a temperature setting that made sense for one task and not another, combined with an evaluation harness that wasn't measuring quite the right thing. Four variables, each individually reasonable, together producing an outcome that makes no sense until you've traced all of them simultaneously.
Junior engineers - and this is not a criticism, it's just a stage - tend to look for the one thing. The bug. The root cause. The fix. And often, for well-scoped problems, that's exactly right. But AI systems at scale don't usually fail that way. They fail the way complex systems fail: through the interaction of components that each look fine in isolation.
This is exactly what you're looking at when you look at the night sky.
Individual stars don't tell you much. A single point of light is just a point of light. But the moment you start seeing constellations - the moment your brain overlays a framework of spatial relationships onto what otherwise looks like noise - the sky starts to communicate. Not because the stars changed, but because your relationship to them did. You developed an eye for arrangement. You brought a model to the observation.
There's a direct line between learning to read a page layout and learning to read a broken AI pipeline. In both cases, the skill isn't knowing what each element does in isolation - any junior designer knows what a pull quote is, any junior engineer knows what a vector database does. The skill is understanding how elements interact, where the tension is, what the composition is actually doing as a whole. The eye for ordered arrangement is the same eye, whether it's trained on InDesign grids or system architecture diagrams.
That's what debugging AI systems actually requires. Not a faster search through logs. Not a better error message. A mental model sophisticated enough that what looks like random floating dots starts to resolve into a pattern that has causal structure. The ability to look at apparent chaos and ask: what would make this make sense?
I wanted the site itself to embody this idea. The background isn't just stars for aesthetic reasons - it's stars because I genuinely believe that the most interesting problems in AI systems look like that. Distributed, non-obvious, requiring a kind of perceptual patience that most tools don't teach you and most curricula don't reward. The cosmos theme is the most honest thing I could put on my homepage, because it mirrors how I actually think about the problems I work on.
For a long time the night sky on this site was a backdrop - a mood, a metaphor gestured at from behind the content. Then I built the Impact section, and the metaphor stopped being scenery and became the argument itself.
Every quantifiable result I've delivered - the 7× throughput, the $1.3M in annual compute saved, the sub-300ms latency held at peak - is a star. The metrics are grouped into categories, and each category is anchored to a vertex of a heptagon. From each anchor the numbers fan out along faint arms of light - trails that say these points belong together, the way a constellation's lines turn scattered dots into a figure. Hover a star and a readout resolves at the centre of the shape: the value, the project, the client, one line of context. The number was always there. What changes is your relationship to it the moment you reach for it - the same move I described earlier, where the sky starts to communicate not because the stars changed but because you brought a model to them.
The interesting part - the part that connects to the work rather than just illustrating it - wasn't drawing the dots. It was control. A naive scatter looks like noise; a rigid grid looks like a spreadsheet wearing a costume. I wanted neither. So the arrangement is governed by a few rules that are, quietly, the same rules I'd enforce in a system.
The first is territorial. Each star's arm is length-capped at the exact boundary where it would drift closer to a neighbouring category than to its own - a Voronoi edge. No metric can ever trespass into another category's region, no matter how crowded its own corner gets. That's not a visual nicety; it's bounded context drawn in light. The same discipline that keeps a retrieval service from reaching into the scheduler's state keeps the cost numbers out of the reliability cluster. The picture enforces the property I'd otherwise enforce in code review.
The second is anti-pattern by construction. Within each cluster, a star's distance from its anchor is decorrelated from its angle using the golden ratio, so the points never line up into rows or an accidental spiral. The eye reads the result as found rather than plotted - which is the whole point, because a constellation you believe in is one that looks like it was discovered, not arranged.
The third is that the sky breathes. Every load nudges each star by a few degrees and a fraction of its length - bounded entropy, clamped so it can never violate the first two rules. The arrangement is never quite the same twice. A real night sky isn't a static asset, and neither is this one; come back tomorrow and the figure has shifted, the way the actual sky does between seasons. It quietly rewards a second visit.
I find it fitting that the section carrying my hardest evidence is also the one that commits most fully to the site's central idea. The proof and the metaphor turned out to be the same object - distributed points that mean little alone and resolve into something legible the moment you bring an eye to them.
The stack section used to be paragraphs - careful, detailed prose explaining each capability and the metrics behind it. I deleted almost all of it. With the Impact constellation doing the proving, the toolkit no longer needed to argue; it needed to be legible at a glance.
So it became a field of marks, grouped by capability - languages, backend, real-time and voice, data, AI, cloud - each technology shown as its actual logo rather than a word. A logo is recognised pre-verbally; an engineer scanning the page finds their own stack in the time it takes to read a single label. Sourcing them was its own small act of care - real vector marks, dark logos recoloured so they survive on a black background, related tools collapsed into one tile so a category reads as a family rather than a list. The same instinct as the constellation: arrangement that respects the difference between density and noise.
And the logos did something I didn't fully plan for: they brought colour back. After pages of near monochrome - white type on a near-black void - the stack is a wash of brand colour, a deliberate exhale. It sits late in the page on purpose, a breath between the weight of the work and the quiet of the close. The constellation taught me the lesson and the stack repeats it: colour, used sparingly, is rest for the eye - and rest is part of rhythm, not the absence of it.
One more thing I want to be honest about: the smoothness.
There's a lot of scroll logic on this site. Layout-effect hooks that measure DOM positions before paint. Intersection observers stacked on top of requestAnimationFrame loops. Sticky header behaviors that required three refactors to get right. Easing curves I tweaked by feel rather than formula.
None of this is visible. That's the point.
The most invisible investment on this site is the scroll itself. The native wheel event is intercepted; the page drives its position through a loop that closes a fraction of the remaining distance on every animation frame. The result is that the page moves with physical weight - not instantly, not with snap, but with the kinetic sense of something that has mass. Most visitors register it as "feeling smooth" without being able to say why. What they're feeling is the gap between input-speed and world-speed - the same gap that makes a flight simulator feel real and a PowerPoint feel like a document.
Smoothness is the thing you only notice when it's absent. When a scroll jank interrupts an animation, when a header snaps instead of slides, when a transition breaks the illusion of physical continuity - that's when the user's brain surfaces from the experience and remembers they're looking at a website. The spell breaks. The immersion ends.
Print designers know this feeling instinctively. A registration error that shifts the cyan plate by half a millimetre ruins an image that might otherwise be invisible. A line that doesn't hang correctly pulls the reader's eye wrong without them knowing why they feel uncomfortable. The flaw doesn't have to be named to be felt.
I think about this the same way I think about latency in AI systems. A response that takes 3 seconds and a response that takes 8 seconds are both "slow" - but the 3-second one might stay in the zone where a user stays engaged, while the 8-second one is the moment they open a new tab. The difference between a system that feels fast and a system that feels broken isn't always about the underlying capability. It's about whether the perceived behavior matches the user's mental model of what should happen.
A portfolio that stutters - even slightly - signals something. Not necessarily that the engineer who built it can't write smooth code. But the human brain is pattern-matching constantly, and a site that doesn't feel polished is a site that subtly undermines its own argument.
If I'm telling you I build careful, considered systems, the first evidence I can offer is this one.
Then someone opened the site on their phone and it didn't scroll.
Not slowly, not jerkily. It loaded - the first screen sat there, composed and still - and then refused to move. A second phone, the same model, scrolled it perfectly. Same build, same browser, two devices that should have behaved identically. The only difference, it turned out, was that one of them was in Low Power Mode.
The cause was the cleverness itself. To keep the first paint fast, the homepage deferred its own work: it waited for the browser to report a quiet moment before bringing itself to life, and it loaded most of its sections on demand rather than all at once. On a rested phone that quiet moment arrives instantly and the page fills out before you've finished reading the first line. On a phone in Low Power Mode there is no quiet moment - the browser hoards every spare cycle - so the page stayed a single screen of content with nothing beneath it to scroll to. The whole design quietly assumed a device with energy to spare.
The fix was humbling in how much it gave back. Instead of asking the browser to assemble the page at runtime, the entire thing is now rendered to plain HTML ahead of time - present, complete, and scrollable before a single line of JavaScript runs. The animation, the intercepted scroll, the constellation: all of it still layers on top, but as enhancement, not as a precondition for the page existing at all. These article pages had always worked exactly this way, quietly, which is precisely why they never had the problem.
It's the same lesson the rest of this page keeps circling, only pointed back at me. Smoothness you can't see includes the smoothness of a page that simply works on a tired phone. The most over-engineered thing on the site was never the comet or the cursor - it was the unspoken assumption that the machine on the other end had anything left to give. The most considered system is the one that asks the least of the person arriving.
The last thing I reworked wasn't a component at all. It was the order.
A page is a sequence, and a sequence is an argument with a shape. Get the order wrong and even good sections read as a fragmented signal - a list of strong things that never adds up to a person. I wanted the opposite: one continuous beam, where each section sets up the one after it and nothing asks the reader to double back.
So it runs: who I am, then how I got here, then - at the crest of that trajectory - the Impact constellation, the payoff the journey was building toward. The numbers raise a question, how?, and the work answers it immediately: the featured systems, then the broader catalogue. Only after you've seen the results and the work that produced them do the recommendations arrive, so the people vouching for me are vouching for something you've already witnessed rather than something you're being asked to take on faith. Then the toolkit, as reference, and finally the way out.
Identity, journey, payoff, work, validation, tools, contact. Read aloud it's almost a sentence. That's the test I held it to - not whether each section was good in isolation, but whether you could fall through the whole thing without the beam ever breaking.
Every portfolio is an argument. Most of them argue with words: here are my skills, here are my projects, here is what I'm good at. That's fine. Words are efficient. Words are legible.
But I wanted to make a different kind of argument - one that a decade of design work taught me is almost always more persuasive than words. I wanted the experience of visiting the site to carry information about how I think, what I value, and what I'm capable of, before a single line of copy is read. I wanted the visual depth to say I care about depth. I wanted the smoothness to say I care about the details that users feel but can't name. I wanted the cosmos to say I live comfortably in complexity, and I know how to find order in it.
The portfolio isn't over-engineered. It's precisely engineered for a goal that most portfolios don't even attempt: to make you feel something before you decide whether to read anything.
There is something at the bottom of the page that I haven't mentioned. Reach it - scroll through everything - and the content blurs, and a message appears: somewhere between the hero and here, you decided to keep going. Thank you for taking the time to know me a little more than you already did. Ten seconds. A countdown ring. Then it's gone.
A visitor who reached the bottom gave their attention to the whole of it. The least the site can do is notice.
Whether it works is something only visitors can answer. But the intent was never decoration.
The intent was always: make them curious enough to look closer.
And then give them something worth finding.