Hey, I’m Omar.
I’m a software engineer
I’m a software engineer, data scientist, data engineer, business analyst, AI engineer, BU alumnus, Questrom MSBA candidate, and your next hire.Work
PODS — Boston · Jan 2026–Present
YW Apparel — New York · Jun–Nov 2025
PortF — London · Feb–Apr 2025
United Nations — New York · May–Aug 2024
Asphalt Green — New York · May 2023–Jan 2024
Sheffield United FC — Sheffield, UK · May–Aug 2021
My Story
Three continents before I finished school. In order:
Projects
Java · Deep reinforcement learning
A two-person Java project with no ML frameworks: a dual-decoder deep Q-network scores every legal move in multi-agent Risk from 41 engineered features. Won 72% of head-to-head games over a 31-game evaluation and matched a 30-rollout MCTS agent with one forward pass.
Python · Flask · React · SQL · AWS
A campus platform for short-term, activity-based groups — study sessions, workouts, pickup games. 50+ active users, semifinals of an $80K venture competition, and agentic group-matching in place of manual coordination.
TypeScript · React Native / Expo · Vercel Edge · Python · OpenCV
A collectible card game on live football data: every player in Europe's top five leagues and the Champions League is a card rated from this season's stats. Forecasts come from a maximum-likelihood Poisson goals model, and a Monte Carlo simulator replays the season 3,000 times.
JavaScript · React · Canvas · Linear algebra
An eight-planet arcade game where every frame is linear algebra — motion, orbits and collisions computed as explicit vector and matrix operations, with no game engine. Each mission runs on its planet's real physical constants.
Python · YOLOv8 · scikit-learn · Flask
Photograph your fridge and get back what's inside, how fresh it is and what to cook. YOLOv8 finds the items, a freshness classifier I trained on 44 hand-built features scores each one (97.5% held-out), and TF-IDF retrieval matches recipes.
Image: Real output on a Wikimedia Commons photo by Candeadly (CC BY 4.0)
Python · NumPy · SciPy · pandas · Flask
Retail price optimization end to end: per-SKU demand elasticities from log-log regression, then profit-optimal prices inside guardrails. On a simulated two-year panel of 43,800 transactions it lifts daily profit 24.6%, and the estimates recover the known ground truth.
Python · pandas · SciPy · MusicBrainz API · Jupyter
A team analysis of what gets a Hot 100 song into the Top 10: an 8-stage Python ETL fuzzy-joins the chart archive, lyrics and MusicBrainz into 5,342 songs. Artist history is the strongest predictor (5.8% → 18.5% Top 10 rate), and we caught a target leak.
React · TypeScript · Vite · Gemini API
A nutrition tracker for BU students that logs meals against what the dining halls are actually serving. Staff post the day's menu from a portal, and Gemini estimates each dish's macros from its name and description.
More builds
Six forecasting models, naive baselines to Holt-Winters and ARIMA, compared under walk-forward backtests with 95% intervals.
A fully offline retrieval-augmented chatbot whose answers cite the exact source sections.
LoRA from scratch in NumPy: a rank-8 adapter (9.6% of the parameters) on a frozen base model, A/B'd side by side.
A from-scratch multi-agent framework — planner, debaters, fact-checker and judge — streamed live to a dashboard.
Version, test and A/B prompts like code, with six local scorers and a running leaderboard.
Toolkit
Elsewhere
The closest thing I have to a system clicking into place — structure underneath, room to move on top.
The same appeal as the analytics work, stripped to its bones: incomplete information, a clock, and no one to blame.
Beat me if you can →
I swam competitively, then taught it — a hundred-odd lessons at Asphalt Green. Football is the other one, close enough that I built an analytics platform for it rather than just watching.
Contact
Always up for a conversation about football analytics, or about what actually happens to a model once someone has to act on it.
