what stayed

I have played squash since I was seven years old in Dubai, which means it has been part of my life longer than any project or company or idea I have worked on. I grew up at a club where the standard was high enough that losing was the default for years and winning had to be earned through incremental work that did not show up in results for a long time. That environment teaches you something specific: you stop looking for the session where everything clicks and start treating improvement as a long compounding problem where the only variable you control is whether you showed up today and worked the right thing.

At Purdue I made the traveling team and we qualified for nationals in the 2025-26 season. I am running for Vice President of the club, with a platform around building the roster, running more internal tournaments, moving to bi-weekly ladder updates, and making practices more structured around drill curriculum rather than just match play.

Squash is a game of position and pressure, not power. The point is usually won three shots before it ends, through movement that forces a bad reply, not through a single brilliant shot. You learn to think several exchanges ahead while staying completely present in the one you are in. You also learn to lose without collapsing, because at the level I grew up playing, you lose constantly for long stretches and the only way through is to treat each loss as information rather than verdict.

That second part is the one that transferred most directly to everything else. The tolerance for losing over extended periods without changing your underlying assessment of whether the work is worth doing is a specific skill and it does not come naturally.

I started cycling to find the edges of a city everyone said was too safe. Ten kilometers became twenty, then sixty. I never planned the route home.

When I got lost — and I got lost often — I'd find the Burj Khalifa on the horizon and navigate from there. You don't need the whole map. You need one landmark you can move toward, and the willingness to correct as you go. I've built almost everything since the same way.

Expo 2020 put a pavilion from nearly every country on earth within cycling distance. I went fifty-five times. Not for the architecture — for the people staffing the pavilions, who would talk to anyone curious enough to ask.

I learned to ask the wrong question on purpose: the one slightly off from what you're supposed to ask, because that's the one that gets you a real answer instead of the brochure.

I started volunteering at Tender Hearts Arena in Dubai in 2021. The center works with children who have special needs, and I taught art and basic Punjabi there for two years, over a hundred hours in total.

The Punjabi teaching started because I speak Punjabi and wanted to do something genuinely useful rather than fill a volunteer hour. I taught fourth and fifth grade students to count in Punjabi by tying the numbers to physical movement: push-up counting, counting while jumping, passing objects while calling out numbers. Abstract language became physical memory. It worked better than drilling vocabulary on paper, and I learned that from watching what actually held the students' attention rather than from any teaching method I had read about.

The harder part of those two years was learning to adapt mid-session. These students had a wide range of attention spans, learning modalities, and communication styles. A lesson plan that worked one week could fail completely the next with the same group. You had to read what was happening in the room and change course while it was still running. That responsiveness is a skill I did not have when I started and built slowly across two years where the feedback was immediate and unambiguous.

I also secured $10,000 in CSR funding for the program. I wrote the proposal, identified the corporate sponsors, and presented the case. It was the first time I had raised money for anything, and the process of turning a real need into a fundable argument and then presenting it to people who had not asked to hear it was something I have used in different forms many times since.

After the physics class moment I had a new working assumption: the cost of speaking was lower than I had been estimating for years. What I did not have was a room that was structurally designed for the kind of conversation I wanted to have, where students from different schools could argue about ideas in front of people they had never met, with no project to present and no winner at the end.

I had noticed that every STEM competition in the UAE was output-focused: bring a project, demonstrate a result, take your ranking home. There was nowhere to just think out loud with other people who were thinking hard about similar things. The gap was not a lack of ambitious students. It was a lack of venue.

I recruited my closest friend first, and he said he had never done anything like this and was too scared to mess it up. I told him I knew exactly how that felt because I had felt it six months earlier. We built it together from nothing: no prior event infrastructure, no venue pre-booked, no list of schools to invite. I handled venue, scheduling, speaker outreach, and participant communications while coordinating a thirty-person organizing team that had never worked together before.

A hundred students came, sixty percent of whom had never left their school for an external event. The structured discussion formats I had designed were deliberate: I knew from my own experience that unstructured rooms default to the loudest voices, and the whole point was to create a room where people who had never spoken publicly could engage without being required to volunteer into silence. Several participants went on to run their own initiatives the following year.

The student who said afterward that it was the first time he had felt free to say what he was actually thinking was not describing a transformation. He was describing the removal of a constraint he had been living inside. The Quantum Quorum was not a STEM event. It was the answer to a question I had been sitting with since ninth grade: what happens if the room exists.

I joined the organizing committee for TEDxYouth Dubai in 2022 as one of the younger members of the team, working on operations: run-of-show materials, speaker prep logistics, cross-team communications, and day-of execution.

The day-of work was where I learned the most. A live multi-stage event has a rhythm that does not accommodate improvisation, and the only way to make it run is to have built enough redundancy into the plan that when something goes wrong, and something always goes wrong, you have a decision ready rather than a scramble. We had technical failures, speakers who ran over, transitions that compressed under pressure, and moments where the schedule was in genuine danger. What I watched the experienced people on the team do in those moments was not panic. They went one level deeper into the plan they had already made and found the slack that had been built in deliberately.

I contributed to the speaker briefing process and helped ensure that talk structure met TEDx guidelines and time constraints. Getting a speaker who has been developing an idea for years to compress it into eighteen minutes without losing the core is its own problem. Doing it in a way that feels collaborative rather than editorial requires a precision in conversation that I had not practiced before.

What running a high-stakes live event deposits in you is a specific awareness of the gap between a plan and what actually happens, and a respect for the people who bridge that gap in real time rather than in retrospect.

We didn't place at the competition where it started. The walk out of that room, the silence between me and my co-founder — the thing that had cost me the most to even try had failed the most publicly. Most teams stopped.

We went back. 300 user interviews, a rebuilt matching algorithm, redesigned onboarding, a dynamic pricing model that closed paid pilots within two weeks of launch. We snuck into career fairs and set up our own desk without permission, because that's where the parents were. I'd spent my whole life waiting to be invited into rooms; that was the first time I just walked in. The room didn't collapse. 700 students and 400 alumni across 15 schools, 60% program completion, a legal license, revenue, a $750 accelerator scholarship, a pitch at GITEX.

A year after not placing, I came back to that same competition as a judge — the only team from that cohort that outlasted it.

It started with a problem and a person, not a brief. Mr. Gopal, my neighbor, is visually impaired and wanted to walk independently. I fabricated the casing from recycled carbon fiber and hemp resin, salvaged a gyroscope from an old smartphone and a motor from a toy car, wired ultrasonic and IR sensors to an Arduino, laid out a custom PCB for sensor power distribution and signal routing, and wrote the obstacle-detection firmware from scratch.

I tuned the vibration feedback through repeated blindfolded hallway testing, and iterated four physical prototypes before it was reliable enough to hand to someone who actually depended on it.

I prototyped a VTOL airframe through many iterative FDM cycles, choosing composites and running structural analysis against thrust-to-weight. The motor mounts and landing gear were modular on purpose: a field swap should not need a toolbox.

I sized the propulsion stack from first principles, power and load by hand before trusting any datasheet, then cross-referenced the ESC, motor, and battery for electrical compatibility. I wired and bench-tested the full stack and logged every result and failure mode before anything went toward flight.

I kept design-iteration logs and wrote down why each material was chosen, so the next person could reproduce the work instead of guessing.

I ran over seventy esports events across the MENA region: more than ten thousand players across several games and formats. I managed a fifty-person operations team covering scheduling, broadcast, referees, and on-ground logistics.

I wrote the run-of-show templates, the rulebooks, and the ops playbooks we reused event to event, which cut setup time hard. The part that taught me the most was live incident management: a stream dies, a match is disputed, a schedule slips, and a few thousand people are watching. You fix it calmly while it is on fire, or it stops being an event.

I was in high school the whole time.

The problem was specific: students sit down to study and pick up their phones within minutes without noticing they did it. Screen time data shows it happening but tells you after the fact. We wanted to catch it in the moment, on-device, without a phone.

The idea was smart glasses with an OV7670 camera running a TFLite distraction detection model at fifteen frames per second, entirely local, no cloud, sub-100ms inference. Daksh, Hema, Amogh and I built it over the course of a semester for the FYE Design Competition at Purdue.

Three prototype generations. G1 was built around the MAX78000 microcontroller because its ultra-low-power neural network accelerator was theoretically the right chip for on-device ML at that power budget. It never shipped. Supply chain failure: the chip was backordered past our timeline and we could not source it. G1 died in hardware before it started. G2 moved to a Raspberry Pi 5 for processing power. The inference ran, but it failed on three criteria simultaneously: too heavy for a glasses form factor, too expensive at roughly $80 in compute alone, and battery life under two hours at continuous inference. We had a working model on hardware nobody would wear.

G3 was the answer. Arduino Nano 33 BLE Sense, OV7670 camera, TFLite model quantized to fit inside 256KB of program memory, under $35 BOM, twelve grams of compute, sub-100ms latency on phone detection, battery life long enough for a full study session. The frames were TR-90 thermoplastic, the same material used in commercial eyewear.

Every decision across all three generations ran through a weighted decision matrix scored on five criteria: ML accuracy, inference speed, wearability, unit cost, and battery life. G1 scored 52.7%. G2 scored 49.1%. G3 scored 69.1%. The matrix was not a post-hoc rationalization. It was the mechanism we used to make calls when two engineers wanted two different things, and it meant that when G2 failed, the failure had a number attached to it rather than a feeling.

I was Lead Product and Systems Engineer. I scoped the system architecture, owned the decision matrix, built the product narrative, and delivered the final pitch. We won $600 at the PESC Design Competition.

What that project taught me was the difference between a prototype that works and a prototype that answers the question you are actually trying to answer. G2 worked. It just answered the wrong question. G3 answered whether someone would wear it and use it, which was the question that mattered.

A.I.D.E. listens for a paste event through the VS Code API and instantly generates a short, targeted quiz on the code you just dropped in. You can't move on until you've engaged with it. It turns passive copy-paste into active understanding.

You define a rule in plain language — if event A resolves true, fire trade B — and the engine watches the order book and executes with sub-50ms detection. A LiveKit voice agent surfaces relevant news and pushes back on your assumptions before anything fires, to slow the impulsive trade.

Built at NexHacks, Carnegie Mellon.

Hack the Future @ Purdue takes about 1.82 percent of applicants. Once you are in, you build real tools for NGOs, not demos. The org I work with serves over 2,000 women globally.

I ship full-stack features in 24 to 72 hour sprints: React and Node and PostgreSQL, role-based access control, dashboards, CSV pipelines. I also wrote Python and OpenCV utilities for automated field-data quality checks, because the data comes from low-connectivity environments where you cannot assume a clean upload or a second try.

The constraint that shapes everything is the network. If a feature only works on good wifi, it does not work.

Humanoid Robotics: I am designing a six-DOF arm and hand in SolidWorks, modular finger and thumb subassemblies for independent bench testing. I replaced tendon-based actuation with rigid linkages and linear actuators, which cut the actuator count and made the failure modes predictable. Flat-profile finger geometry as the fallback when FDM tolerances drift.

IEEE Racing: sensor mounts in Fusion 360 and Siemens NX, part count down fifteen percent through DFM analysis, and an Arduino PID loop for the drivetrain that I tuned gain by gain on the bench and the track to kill overshoot.

Baja SAE: suspension and drivetrain for a single-seat off-road car, GD&T and weldment analysis on the tubular frame joints, stress analysis at the interfaces until the safety factor holds inside the weight budget.

Data centers already draw roughly one percent of the world's electricity, and the densest racks are passing the point where moving air can keep them cool. Two-phase immersion cooling submerges the hardware in a dielectric fluid that boils off the heat directly.

I support experimental design, data collection, and iteration on system performance under varying thermal load, and presented our findings at a Purdue campus expo to a mixed technical and non-technical audience.

Many people with Parkinson's lose the reflex to swallow on schedule. Saliva pools, and aspiration follows — quietly, often. ChYme reads throat and sternum motion with two MPU-6050 IMUs over I2C and runs a four-class CNN on-device to catch a swallow and, when too long passes, remind.

I was the embedded hardware lead. I designed the power circuit from scratch — TP4056 charger, MT3608 boost, a LiPo cell — wired the dual IMUs, and wrote the STM32 ring-buffer firmware feeding the classifier. 93% accuracy running on an Arduino Uno Q.

Around hour twenty the team split. Software wanted maximum accuracy; I wanted something a shaking hand could actually wear. A device that hits 95% but sits wrong on the neck ends up in a drawer — and a device in a drawer has zero accuracy. We stopped arguing specs and asked who we were building for. Both the form factor and the accuracy got better after that.

Three playable industry labs — music production, motorsport strategy, a finance deal room — plus ten global pioneer profiles and a future-letter system. Zero-pressure branching UI so exploration felt like a game, with the Claude API giving each path its own mentor.

I led a four-person team through a 36-hour hackathon: architecture, API design, schema, deploy, the final pitch. It won first out of fifty-plus teams and is still live, still getting traffic.

I work on natural-language coordination protocols for multi-agent human-robot environments, inside the lab's Natural Language Environments framework. The question I keep returning to is where instruction parsing breaks: how it degrades in spatially dynamic, multi-agent contexts, and what fallback architecture is needed when semantic grounding fails in unstructured physical space.

The other thread is agentAR — tool-augmented LLM agents that support in-situ AR creation. Same problem, different surface: an agent has to act inside a physical scene it only half understands.

What I keep circling is robustness. A plan that survives contact with a clean demo is not the interesting case. The interesting case is the agent two steps off the script, improvising with whatever it has left.

A tire is loud. All that vibration and acoustic energy is normally wasted. We are building a metamaterial-enhanced harvester that captures it across the frequency spectrum of a rolling tire and uses it to power low-energy wireless sensors, for tire pressure monitoring and road condition sensing.

I own the mechanical side: the structural design, where the transducer sits, and how it couples acoustically to the tire. My partner handles the PCB and the energy-harvesting IC. I am reading my way through equivalent-circuit models for piezoelectric harvesters and the primary literature on sound-field control, because the geometry decides what the circuit ever gets to see.

Founders with high ad spend and low conversion think the product is broken. Usually it's the website — and they can't see it, because they know it too well.

Candid deploys fifty Playwright browser agents on Modal, each a different user with its own viewport, accessibility settings, and behavioral script. The full sweep finishes in under sixty seconds. A smart-click routine with twelve fallback strategies and a UX-versus-tool failure classifier keep the scores honest. It runs axe-core, Lighthouse, and a custom AI-SEO auditor that checks your robots.txt against a dozen AI crawlers. Gemini 2.5 Pro synthesizes all fifty sessions into one report — ranked fixes, annotated screenshots, a one-click prompt for the code-level solution. Supermemory indexes each run by industry, so every audit is sharper than the last.

I built it solo at HackIllinois after my team dropped out fifteen hours in. A week later the same idea won the JHMC Ideathon and a thousand dollars: there I defined the ICP, the jobs-to-be-done framing, and pricing at nineteen dollars a test or ninety-nine a month. The thousand went to a one-way flight to San Francisco.

The Ideathon prize money went straight into a one-way ticket to San Francisco. No housing sorted, no meetings lined up — just a goal I'd set at the start of the year: to be in rooms with people building things in person instead of through a screen.

A week of late nights and long conversations with founders and investors I had no real business being around yet. Research leads, internship talks, advisor offers from people I'd only read about. The thing I hadn't been chasing showed up the moment I stopped building for it.

I wrote this essay for a Purdue writing competition in early 2026 and won first place. The essay used Plato's Ring of Gyges as the lens for a question I had been sitting with without naming it.

The Ring of Gyges story comes from Book 2 of the Republic. A shepherd finds a ring that makes him invisible, and within a short time uses it to murder the king, seduce the queen, and take the throne. Glaucon's argument is blunt: anyone would do the same. Justice is not something we choose because we value it. It is something we perform because we are watched. Remove the audience and there is no performance.

I read this and felt something uncomfortable settle in my chest, not because I disagreed but because I was not sure I did.

The question the essay was built around: if no one could see the wins, would I still chase them? If the resume disappeared, if the LinkedIn post never got written, if the award existed only as a memory I could not share, would the work still feel worth doing? I sat with that longer than I expected. The honest answer, the one I did not want to give, was that I was not sure.

Socrates' response to Glaucon is not to lecture him. He asks what kind of person Gyges became in the process. The unjust person is not free, Socrates argues. He is enslaved to appetite and status-seeking in a precise technical sense: his reason no longer governs him. The tyrant, the ultimate Ring-of-Gyges winner, lives surrounded by enemies, constantly reactive, unable to trust anyone. No one is more wretched.

What the essay landed on was a distinction I have come back to many times since: the difference between building something and performing the act of building something. What would survive if no one was watching is the actual work. What would fall away is the story I tell about the work. The essay was honest that those two things are harder to separate than I would like them to be, and that the ambition I carry is not cleanly one or the other. It is both, and the work is in knowing which is driving at any given moment.