Robotics & AI Engineering

Hussain
Sabuwala

BEng Robotics & AI, University of Birmingham Dubai

View projects ↓Get in touch

PHASE 02 · LOCOMOTION

About

hussain@hs01:~$

> System_Status:

I'm a Robotics & AI engineering student at the University of Birmingham Dubai. I like building things that connect software, hardware, data, and real-world engineering problems.

Most of my work sits somewhere between research and building. I've worked on satellite-data pipelines for urban heat analysis, machine learning projects, robotics and systems engineering, and full-stack products for hackathons and personal projects. I'm also currently working on my first research manuscript through my NYU Abu Dhabi urban heat research, where I've been analysing Dubai's heat patterns using satellite data, geospatial pipelines, and statistical checks.

I'm still early in my journey, but I care about building things properly. I like projects where the result is not just a demo, but something that has structure behind it: clean code, testing, useful data, engineering reasoning, and a clear problem it is trying to solve.

PHASE 03 · POWER

Builds

BUILD 01

LIVE · EARLY ACCESS

Jacobi

A tool that catches websites charging you more — and proves it with statistics.

Jacobi dispatches 24 synthetic shopper identities against a product page — varying location, device, cookies, referrer, and language — then attributes any price gaps to their causes. The 0–100 exploitation score hard-gates to zero unless the statistics survive a coverage gate: it refuses to claim discrimination it can't prove. Live on an early-access waitlist with a Chrome extension and research-grade PDF evidence reports.

  1. 01Per-variable price attribution computed as a finite-difference Jacobian — the name is the method
  2. 02Welch's t-tests behind a coverage gate, so the score collapses to 0 without significance
  3. 03Multi-network egress (datacenter / residential / mobile) to separate network effects from real discrimination
Source ↗Live ↗

Python · FastAPI · NumPy · Next.js · Supabase · Bright Data

BUILD 02

RESEARCH · MANUSCRIPT IN PREP

Dubai Urban Heat Stress

The satellite pipeline behind a journal manuscript on Dubai's urban heat.

Built during my NYU Abu Dhabi research internship under Dr. Rita L. Sousa: 15 months of Landsat 8/9 and ERA5-Land data processed into 90 monthly layers, with zonal statistics across 75+ urban, vegetation, desert, and water sample zones. The analysis produces pixel-level hotspot persistence maps and a prototype weighted Heat Exposure Index, all feeding a manuscript currently in preparation. The public repo stays deliberately high-level while the paper is in review.

  1. 01Landsat surface temperature fused with ERA5-Land reanalysis, so surface and atmosphere tell one story
  2. 02Class-based zonal statistics instead of city-wide averages — heat exposure differs by land type
  3. 03Uncertainty carried through every figure: 95% confidence intervals and sensitivity checks before any claim
Source ↗

Google Earth Engine · QGIS · Python · Landsat 8/9 · ERA5-Land · Streamlit

BUILD 03

PROTOTYPE

Agent Sanad

The AI reads. Deterministic code decides. A human owns the exceptions.

A decision-support prototype for UAE housing-loan arrears rescheduling, built for the Agentera hackathon (MOEI × 42 Abu Dhabi). The language model only reads and summarizes evidence — every decision comes from a deterministic policy engine, so the system stays transparent and auditable instead of a black box. It turns a five-working-day manual review into a sub-second, fully audited draft.

  1. 01A deterministic policy engine owns every outcome; the LLM is never in the decision loop
  2. 02Benchmarked against held-out 2025 cases before claiming any accuracy
  3. 03Append-only audit trail with tamper-detection receipts, ready for a real compliance review
Source ↗Live ↗

Python · FastAPI · Policy Engine · Pytest

BUILD 04

IN DEVELOPMENT · RESEARCH

CardioFlowX

From hackathon finalist to research pipeline: teaching a machine to see coronary arteries.

CardioFlow started as a national-hackathon prototype — a ~10M-parameter physics-informed network over a 1D Navier–Stokes solver, demoing FFR-style estimates in under 0.75 seconds on a laptop CPU; we were shortlisted to the final 10 of 100+ teams. It's now being rebuilt as CardioFlowX, a research-grade pipeline: a patch-based 3D U-Net trained on real CCTA volumes (ImageCAS) for coronary artery segmentation, centerline and stenosis geometry extraction, and a physics-informed hemodynamic surrogate on top. A research prototype, not a clinical tool — and honest about the difference.

  1. 01Segmentation → centerline graphs → physics-informed flow, one validated milestone at a time
  2. 02Patient-level splits and vessel-positive patch sampling before any accuracy claim
  3. 03Navier–Stokes residuals inside the loss — physics enforced, not implied
Source ↗Live ↗

Python · PyTorch · 3D U-Net · PINNs · Navier-Stokes · DICOM

BUILD 05

COMPLETE

Project Urban

Forecasting Dubai's 2027 urban footprint from seven years of Sentinel-2 imagery.

An end-to-end geospatial ML pipeline across 155,000+ hectares: engineered NDVI/NDBI/MNDWI and terrain features, a 50-tree Random Forest forecasting ~2,936 hectares of new development by 2027, and a fine-tuned ResNet-18 auditing the training labels at 95.5% validation accuracy. Documented end to end on GitHub so anyone can reproduce it.

155k+ ha area covered · 95.5% label-audit accuracy

Source ↗

Python · PyTorch · Google Earth Engine · Pandas · Random Forest

BUILD 06

HACKATHON

FluxGate AI

A digital-twin command center for school traffic, built in one weekend.

48 hours with a small team: a real-time 2D/3D digital twin of school zones and bus bays, with AI arrival prediction, traffic-density-triggered gate control, and RFID/NOL vehicle verification. Placed top 5 nationally.

48 h build time · Top 5 national placement

Source ↗

React 18 · TypeScript · Vite · Tailwind

BUILD 07

COMPLETE

Odoominance

A Dubai-style food-safety inspection engine, native to Odoo 19.

Models a municipal inspection end to end: a seven-state workflow from draft through supervisor sign-off, weighted checklists where every item carries a severity multiplier, and auto-computed compliance percentages, risk bands, and A–F grades — with role-based access, demo data, and integration tests.

7-state inspection workflow · A–F computed grading

Source ↗

Odoo · Python · PostgreSQL · XML

PHASE 04 · COMPUTE

Field Log

  1. MAR 2026 – PRESENT · ABU DHABI, UAE

    JOURNAL MANUSCRIPT · IN PREP · 2026-07

    NYU Abu Dhabi

    GIS & Remote Sensing Research Intern

    90

    monthly layers

    75+

    sample zones

    95%

    confidence intervals

    • Working under Dr. Rita L. Sousa and Dr. Basam Dahy on Dubai's urban heat, with the work going into a journal manuscript currently in preparation.
    • Built the full data pipeline: 15 months of Landsat 8/9 and ERA5-Land data turned into 90 monthly layers, with zonal statistics across 75+ sample areas covering urban, vegetation, desert, and water.
    • Ran uncertainty analysis with 95% confidence intervals and sensitivity checks, plus pixel-level hotspot persistence maps and a prototype weighted Heat Exposure Index.

  2. SEP 2025 – PRESENT · DUBAI, UAE

    UoB Rocket Propulsion Labs

    System Engineer · site ↗

    • Part of a 10-person student team representing the University of Birmingham Dubai at the UKSEDS National Rocketry Championships 2026 and Mach-X.
    • System-level design, cross-subsystem integration, and launch-readiness planning for the competition rocket.

  3. JAN – APR 2024 · TORONTO, CANADA

    University of Toronto · ESP II

    Research Coordinator

    Scored 100 smart-home system combinations with Pugh matrices and weighted MCDA for PDR Technologies; co-wrote the 70-page design specification.

  4. SEP – DEC 2023 · TORONTO, CANADA

    University of Toronto · ESP I

    Research Coordinator

    Drafted the design constraints for Hart House's North Alley redesign and helped narrow 100+ concepts to three finalists.

PHASE 05 · PERCEPTION & MANIPULATION

Capability Matrix

SYS-LNG

Languages

// what I use most

  • Python●
  • TypeScript / JavaScript●
  • C / C++
  • SQL
  • MATLAB

SYS-ML

ML & Statistics

// models with proper evaluation

  • PyTorch●
  • CNNs / ResNet transfer learning
  • Random Forest
  • Hypothesis testing (Welch's t)
  • Model evaluation & uncertainty

SYS-GEO

Geospatial

// the research toolchain

  • Google Earth Engine●
  • QGIS
  • Sentinel-2 / Landsat 8-9
  • ERA5-Land reanalysis
  • Zonal statistics

SYS-DATA

Data Engineering

// wrangling and visualising data

  • NumPy / Pandas●
  • SciPy
  • Streamlit
  • Tableau

SYS-WEB

Backend & Web

// turning projects into usable products

  • FastAPI●
  • Next.js / React●
  • Supabase / PostgreSQL
  • REST APIs
  • Vercel / Render / Docker

SYS-SYS

Systems Engineering

// from the rocketry team

  • Requirements & trade studies
  • Pugh matrix / MCDA scoring
  • Cross-subsystem integration
  • Launch-readiness planning

● The dot marks the tools I reach for most. I'm still early and learning a lot, and right now I'm picking up ROS 2, control theory, and embedded C++.

Model registry

// models I've actually trained, with their real numbers

  • resnet18-label-audit

    CNN · transfer learning

    95.5% val accuracy

    Sentinel-2 training-label auditing · Project Urban

  • rf-urban-growth

    Random Forest · 50 trees

    2,936 ha forecast

    Dubai 2027 expansion · Project Urban

  • unet3d-coronary

    3D U-Net · patch-based

    in training

    CCTA coronary segmentation · CardioFlowX (ImageCAS)

  • pinn-hemodynamics

    physics-informed surrogate

    prototype

    1D Navier–Stokes pressure/flow · CardioFlowX

PHASE 06 · ACCEPTANCE

Get in touch

ALL SYSTEMS NOMINAL // HS-01 ACCEPTED

I'm open to research, internships, and problems worth building for. Tell me what you're working on — I read everything that lands here.

Hussain Sabuwala

Operator

Hussain Sabuwala

Usually building. Replies faster than the robot assembles.

Field test · proving ground

00000