Stories & Projects | Umar Qureshi

Stories & Specs

I’ve always hoped to build things people actually want to use. This section brings together some of the projects I’ve worked on, the ones in progress, and a few I’ve just been curious enough to mess with.

Some are polished systems. Some are experiments. All have stories behind them, along with specs if you’re into that.

Scai™ — My First AI Assistant

I’ve always been fascinated by the idea of a digital AI assistant. Something like Jarvis from Iron Man, or TESS from Salvation. Not just a chatbot, but an intelligent interface that understands you, represents you, and works alongside you.

Scai (short for Smart Conversational Augmented Interface™, pronounced Sky) is my first attempt at building something like that.

It’s a conversational layer that lives on this site. Always on, always ready. It answers questions, shares context, gives me a nudge when needed, and hopefully helps connect me with future collaborators. One step closer to building something that feels like a second brain.

I trained it on my background, projects, goals, and the kinds of things people usually ask me: what I’ve worked on, what I’m exploring next, what I care about building. Instead of reading through a resume or scrolling LinkedIn, you can just ask.

Under the hood, it’s a multi-turn, retrieval-augmented service. It remembers the flow of conversation, is stateful, and built to reduce friction. Powered by Gemini, running on Google Cloud infrastructure. Lightweight, persistent, and quietly doing its thing in the background.

I didn’t ask Scai to write this.
…Or did it?

Human Motion Meets Deep Learning

I spent the last several years working at the intersection of biomechanics and machine learning, building systems that understand how humans move.

It started with wearable sensor data: raw time-series streams of motion. I designed a Gait Understanding Engine, a deep model that segmented that data into meaningful gait phases, making it possible to extract and analyze real-world movement patterns.

Training these models hasn’t always been easy. The data is messy, varies across individuals, and can be hard to label. But that’s what makes it interesting. Moving forward, I built models to classify different movement patterns in the wild, trained machine learning models to predict and correct user errors, and developed statistical algorithms including an adaptive quaternion-based Kalman filter to estimate orientation, speed, and distance while correcting for drift and noise.

I built AI-powered conversational agents to help users understand what the system was telling them. These agents guided users through the product, answered questions, and helped them troubleshoot issues on their own.

But it wasn’t just about models. I built the systems around them — data processing pipelines to structure sensor streams, tools to catch edge cases, retraining loops to handle distribution shifts in production, and the infrastructure to turn ideas into real, deployable systems. These became essential to the product, powering insights, unlocking features, and driving real business value.

Different problem, same goals: reduce friction, make complex things simple, and build tools people actually want to use.

From Models to Meaning: Seeing the Bigger Picture

My journey started with building models — training them, tuning them, validating the results. But it didn’t take long to learn the model is just one piece. Making the whole thing actually work is where it gets interesting.

In startup environments, sometimes there’s no clear map. Just a rough idea and a problem that needs solving. Building from scratch gives you perspective. It’s not just about writing code and training models; it’s about designing a system. It leads to better questions — not just “how do we build this?” but “should we?”

Infrastructure is kind of like a puzzle. All the pieces need to fit together. Training workflows, deployment pipelines, monitoring systems. They have to scale, support iteration, and hold up when things move fast. The goal isn’t just to build, but to build well.

So what makes it all actually work?

It’s seeing the bigger picture: translating rough business goals into real systems, collaborating across teams, and building with intent. Sometimes that means designing a smart solution. Other times, it means knowing when a model isn’t needed at all.

Exploring Latent Space

I’d worked with encoder–decoder models before — even trained a simple GAN for fun. But what really pulled me in was how these models learn to represent high-level concepts in abstract ways. Smile. Age. Style. Not as attributes, but as directions. Shifts you could apply, amplify, remix

So I got curious and started exploring. Tinkering with pre-trained StyleGANs, editing tools like InterfaceGAN and GANSpace, and poking around the latent layers to see how tweaks in that space changed what came out the other side.

These models encode features, extract structure from noise, and generate variations that feel oddly familiar — sometimes eerily so.

There’s something poetic about it: a machine dreaming up thousands of ways a face might smile, or how a car design might evolve into something sleeker, stranger, or just… different. It’s not how we think — but it’s not that far off, either.

I haven’t built anything serious with it yet. Not for lack of interest — just time. But I still think about it. How features become geometry. How randomness becomes meaning. How a latent vector can almost feel like language.

It started with a beach volleyball club I ran in Vancouver. What began as a way to bring people together every weekend grew into a full-on community. Fun, welcoming, but also surprisingly time-consuming to manage.

Signups. Coordination. Waitlists. Messaging. Skill balancing. The more the club grew, the harder it got to keep things smooth and fair. Especially when you’re trying to make sure everyone gets to play and actually enjoy it.

So I started building something to fix that.

A social sports community management app. Lightweight, simple, and built for real-world organizers. Not just tools for schedules and attendance, but smart support for managing group sizes, matching skill levels, and keeping the whole thing flowing without constant manual effort.

It’s still in the works, but the idea is simple:

Let organizers organize less, and help everyone have a good time.

Real-World Connections Platform

This one’s still in stealth mode, but here’s the broad idea:

I’m building something for people who want to connect in person, without the noise, pressure, or awkwardness of most digital platforms.

It’s not about swiping, tracking, or chasing. It’s about something quieter. Maybe a bit more human.

The goal? Make real-world connection feel more natural, intentional, and just a little easier.