Saad Muzammil
I build production-grade agentic AI systems, not demos.
Ready for work
Saad is ready for their next project!
The before/after storefront structure captures the actual business model: a small business with no online presence gets a website built for free, and the ongoing relationship is the hosting revenue. That's why the arrow loops back to the laptop instead of just pointing left to right, it's meant to signal a continuing relationship rather than a one-off project delivery, which is the part of the business model that makes it sustainable.
0
5
This one is structured around a single branching line rather than three separate scenes, since the three channels (speaks, reacts, tech) share one creator and one brand identity rather than being unrelated projects. The desk and ring-light setting grounds it in the actual production workflow (OBS, minimal editing), and each channel gets its own color and visual motif so the three feel distinct without breaking the sense that they're one connected thing.
0
5
This one is built around a command-center metaphor because that's functionally what an agentic triage system does, watches many streams, routes tickets, flags anomalies. The dashboard wall represents the fleet management platform itself (routes, trucks, live data). The small glowing "agent" orbs moving along connector lines are the LLM-driven agents doing the routing work in the background, and the one with a red pulse specifically calls out the fraud detection piece as a distinct, higher-stakes function within the same system.
0
4
A master's thesis titled "Global Optimization for PV-Integrated Smart Distribution Grids Using Data-Driven Neural Networks," supervised by Dr. Raheel Zafar and evaluated by Dr. Nauman Zafar Butt. The core idea is a teacher-student setup: an SOCP relaxation acts as a convex "teacher" that produces near-optimal solutions to a voltage-var optimization (VVO) problem, and a feedforward neural network (FCNN) is trained as a fast "student" surrogate that mimics the teacher at inference speed. A constraint-embedding trick using a tanh-bounded output layer cut infeasibility rates from 76.61% down to 0.13%. Validated on the IEEE 33-bus system using OpenDSS, with cyclical time encoding to capture daily/seasonal patterns in the grid data.
0
5