Gun goes off
Computer science at Georgia Tech, 3.83 GPA, class of 2028. Machine learning research at CMU, products with paying customers, and a habit of picking the harder option.
Settle in.
The first miles are the ones nobody claps for.
Where the base miles come from
Moving to a country I had never lived in. Working out how to pay for it. Data structures at 2am, linear algebra the morning after, proofs on a whiteboard until they finally made sense. None of it was clever. It was showing up on the days I did not want to, which is the only way base miles ever get built.
Summer 2025
Computer vision that went into a live retail chain
- Built a real-time computer vision pipeline in Python with OpenCV and MediaPipe, running facial recognition on live CCTV feeds.
- Deployed an inference model for gender and employee classification, automating analysis across 4 concurrent camera feeds in production.
- The output fed the analytics dashboards below — footcount, hourly traffic, and payment behaviour across every outlet.
Keep going.
The long, steady stretch
Xu Lab, Carnegie Mellon University
- Designed and evaluated prompting strategies — bounding box versus text — for MedicalSAM3, a multimodal deep learning segmentation model.
- Analyzed model failure cases across CT and MRI imaging, identifying systematic error patterns and implementing mutual-information-based corrections.
WeLab — Georgia Tech, Human-Centered Computing
- Synthesized peer-reviewed research on multisensory environments into evidence-based design frameworks for therapeutic spaces supporting older adult well-being.
- Evaluated and iterated on technology-supported community solutions through UX research and usability testing.
NES Lab — University of Georgia
- Researching advanced feature selection techniques, focused on improving Beam Search algorithms within ScalaTion.
- Redesigning the search from a greedy, forward-selection approach into a true parallel beam search, so multiple feature subsets are explored simultaneously.
- Optimizing model performance against metrics such as sMAPE, and evaluating improvements across multiple regression datasets.
Halfway.
This is where the race actually starts.
Summer 2026
Integrated Solutions
- Contributed to ISConnect, a contractor management web platform — feature definition, testing, and user-facing improvements alongside the product team.
- Researched, compiled, and validated a structured dataset of businesses for the EV Greenbook application.
- Joined technical and partnership discussions with Greenlink Analytics and Alabama Power, and produced materials for a major conference.
Things built off the clock
LazyChat
An AI sales agent that talks to customers, takes orders, and closes sales across Facebook, Instagram, and WhatsApp. Built on Meta's official messaging APIs and tuned specifically for Facebook replies, where most of these businesses meet their customers. It answers in Bangla, English, and the Banglish mix people really type.
AUTHORISED BY 
· OPTIMISED FOR FACEBOOK REPLIES
BudgetBondhu
Bangladesh's first financial literacy edtech startup. A full-stack prototype for gamified budgeting lessons, plus a curriculum covering savings, budgeting, and financial planning — the things nobody taught us at school.
AI @ Georgia Tech
The club site and its traffic — projects, events, and a blog, running as the front door for AI ATL. Real users, real analytics, and a bounce rate I keep trying to push down.
AquaRoute
AI-driven maritime route optimization using KDTree spatial queries and a modified A* search, visualizing real-time ocean plastic concentration data.
Mile 20. The wall.
Everyone hits it. The only question is what you do next.
Mile by mile
The other training log
The running is not a metaphor I borrowed. I am training for an Ironman, though what I actually want is to get faster on foot. Halfs and fulls are the goal. The bike and the swim are cross training.
Some things you only find out by actually getting to know me.
And the rest of it
Six thousand miles from where I grew up, a kitchen in Atlanta, and people who will argue at length about the correct amount of ketchup. Every training plan has a recovery week built into it. This one is mine, and the rest of the plan only works because of it.
Achievements unlocked
Quick answers
Machine learning and software engineering internships for Summer 2027 — ideally somewhere I'd be shipping to real users, not just prototyping.
LazyChat. It went from an idea to 113 paying clients and $120K ARR, and the hardest part wasn't the GPT-4 integration — it was making it reply naturally in Bangla, English, and the mix of both people actually type.
Medical image segmentation with MedicalSAM3 at CMU's Xu Lab. I compare how the model responds to different prompt types, then read through its failures across CT and MRI scans to find patterns a metric alone won't show you.
Really run. Ironman training at the moment, though the goal is getting faster on foot. The rest you would have to ask me about.
Atlanta, at Georgia Tech.
That's the route so far
The route isn't finished — this is just where the split sits today. If you're building something hard, I'd like to hear about it.