مها العمري AI · University of Jeddah · graduating 2026/27

Maha Alamri

I teach machines to read building codes

I build AI systems end to end. The model is one part. The data, the structure around it, how it is tested and the software that puts it in front of people decide whether it works.

Ask me anything /
learn a real network, training right now

Teach it something. Watch it learn.

A small neural network (2 → 16 → 16 → 1, tanh, Adam), training live. The colour behind the points is what it believes right now. Pick a pattern or tap to add your own points.

0steps
–loss
–accuracy
Loss over time
The network · edge colour is weight sign
data what the model is allowed to know

A score is only as honest as the split behind it.

Two tabular projects where most of the thinking happened before training: deciding what the model may see, and how to test it so the number means something.

Credit Risk Intelligence

project

Loan outcome prediction, compared across TabPFN, Random Forest and Logistic Regression. My first step was removing every field that only exists after a loan has already been repaid or written off. Those fields make a model look brilliant in the notebook and useless on the day it meets a new applicant.

Preprocessing lives in one ColumnTransformer: imputation, scaling and one-hot encoding. Class imbalance is handled with random oversampling, hyperparameters with GridSearchCV, and the decision threshold is tuned for macro F1 instead of sitting at the default 0.5.

Leakage check

Recorded before the decision

annual_incomeloan_amounttermemployment_lengthcredit_history_length
decision

Recorded after the outcome

total_payment_receivedrecoverieslast_payment_datecollection_fees

Flood Risk Intelligence

project

Flood events cluster in space and time. If neighbouring locations land on both sides of a random train/test split, the model can memorise the neighbourhood and the score overstates what it would do somewhere new. So I held out whole regions and time periods instead.

I ruled out SMOTE for the same reason. It creates synthetic positives by interpolating between existing ones, which here means inventing places and moments that never existed.

Same map, two ways to hold out 24 cells
traintest, touches a train celltest, isolated

architecture shape the model to the data

Every column becomes a token.

The flood model is an FT-Transformer: tabular features are embedded one by one, a learned CLS token gathers them through attention, and a small head reads the result.

latitudelongituderainfallelevationslopedistancetime + engineered spatial+ engineered temporal+ interactions
0.8638 PR-AUC on held-out regions and periods. Precision–recall is the curve to watch when positives are rare, and SHAP values show which features drove each prediction.
reasoning language, retrieval, rules, agents

Read the drawing. Read the code. Then decide.

Retrieval gives a language model something true to stand on. Rules decide what must hold. Agents split a job into steps, with a person signing off at the end.

MI'YAR

معيارproject

Mi'yar means standard. It is an assistant for the Saudi Building Code, and the problem is that compliance lives in two very different materials: architectural drawings and long regulatory text, in Arabic and English. MI'YAR reads both, then hands the comparison to a rule engine instead of leaving it to a language model's judgement.

Two inputs, one verdict · select a stage

AgriAssist

real client

I built it for a real client: a conversational assistant for an agricultural company, grounded in its own product catalog and support knowledge. Documents are embedded with OpenAI embeddings into ChromaDB and retrieved through LangChain. Part of the source material was Chinese, so the assistant also handles translation into English.

RAGLangChainChromaDBOpenAI EmbeddingsZH → EN

CareerMatch

live product3 agents

Built with Sarah Alshaikh and running live. CareerMatch turns a CV into job applications. Three AI agents split the work, and a person approves every application before it leaves.

The first agent reads the CV against a target role and returns a match score, matched and missing skills, and specific improvements. The second searches live job listings and ranks them by fit. The third drafts the subject line and cover letter for the chosen job. Nothing is sent until the user approves, edits or cancels it.

Agent handoff
Agent 1CV analyst
Agent 2Job scout
Agent 3Application writer
HumanApproval

Try it live with your own CV.

perception pixels into decisions

Teaching models where to look.

Vision work across a reef, a runway show and a road. Each one asks what the picture has to tell you before anything acts on it.

Environmental intelligence · with a KAUST lab

RedSea-Lens

Coral bleaching detection with deep learning, built with a research lab at KAUST. My contributions include AI-based tracking of coral condition, so a reef's health becomes something you can follow instead of a single snapshot. Jeddah sits on the Red Sea, and a reef changes long before anyone on shore notices.

KAUST collaborationCoral health trackingComputer vision
Pose estimation · deployed at JFW

Virtual Try-On

Built and used as an interactive experience for visitors at Jeddah Fashion Week (JFW). MediaPipe BlazePose finds body landmarks from the camera, and the garment is fitted to that skeleton live, so people could walk up and try pieces on.

Jeddah Fashion WeekLive visitorsMediaPipe
Guardian · road + driver, watched together

Guardian Safety Car

A safety co-pilot that watches two things at once: the road ahead and the person driving. Perception measures how fast each obstacle is closing in. A decision layer then escalates one step at a time, from a quiet notice to a warning to assisted braking, and only takes over when the driver hasn't responded. It is designed to back the driver up, not replace them.

MonitorInformWarnAssist braking
80km/h
–nearest gap, m
–time to collision
Watchingdriver
  1. Switch the driver to “looking away”, then make the car ahead brake.
engineering 28 Jun → 20 Aug 2026

Real systems have other systems inside them.

Summer internship at Prince Sultan Aviation Academy (PSAA), Technical Services.

I worked on this system from day one. I designed and built the database, then connected four departments into one smart, interactive dashboard: maintenance-management data, SharePoint, Simorg and budget figures, kept in sync behind a FastAPI, SQLAlchemy and PostgreSQL backend, with an AI assistant that answers questions over the same data.

Beyond the dashboard, I automated parts of the departments' day-to-day processes, so work that used to be gathered by hand now updates itself. Budget data sits behind access control, because it has owners and not everyone sees it.

How the dashboard is wired · view as
Sources
Maintenance management datarequests · status · specialists
SharePointintegration
Simorgintegration
Budgetsynchronisation
Backend
FastAPIAPI · auth
SQLAlchemymodels
PostgreSQLone source of truth
Dashboard · React
Service requests & status
Specialists
Reliability & quality
🔒 locked
Initial budget · expenditures
AI assistantquestions over live data

Signed in as a specialist. Budget figures stay server-side.

Turnaround Guardian

projectaviation safety

During a turnaround, a fuel truck, a baggage loader, a catering truck and a pushback tug all work around one aircraft against the clock. The danger is timing: a vehicle can be in exactly the right place at the wrong moment.

The goal: reason about ramp risk through the aircraft's state. The same spot is safe while the aircraft is parked and dangerous once the beacon is on or an engine starts. Turnaround Guardian combines anomaly detection and predictive monitoring to warn before the state changes, and a human supervisor stays in charge of every call.

Live ramp · aircraft state
active hazard zonebecomes active in the next statevehicle, clear

  1. Switch the aircraft state and watch which vehicles become a risk.
product when the output has mass

Software that has to move a real joint.

Full Stack Robotics Engineering, a two-month program. I designed a robotic arm in Onshape, built it, and drove it from an Arduino. Mechanics, wiring and control code had to agree before anything moved.

Two-joint inverse kinematics · move your pointer over the plate

teaching the test of understanding

If I can't explain it, I don't understand it yet.

At the AI Pioneers Club, Education Department, I write specialised AI content for workshops, bootcamps, courses and awareness sessions. The hard part is deciding what a newcomer can skip and what they can't, then building the explanation around that.

Training
Microsoft Certified: Azure AI FundamentalsCertification
ELEVATEICAIRE × UNESCO × SDAIA × Microsoft · 5 weeks, 26 hours
AI Engineering BootcampSaudi Digital Academy × WeCloudData
Saudi Gen AI HeroGenerative AI training
Generative AI coursesCoursera
Full Stack Robotics Engineering2 months · robotic arm, Arduino, Onshape
constellation projects and the tools they used

Tools are only interesting in use.

Every project pulls its tools into orbit. Drag a star, or hover to trace it.

Hover a project or a tool.

Also in the kitC, Java, XGBoost, PyTorch, TensorFlow, OpenCV, Tesseract, PaddleOCR, ezdxf, FAISS, Sentence Transformers, Pandas, NumPy, Matplotlib, Plotly, Power Query, Talend, Informatica, Express, GitHub, Docker, Nginx, Azure, AWS basics, n8n.

now in progress and open questions

What I'm working on.

My graduation project, and the questions I keep coming back to.

Dark Fleet

graduation project · in progress

Some vessels switch off their tracking signal and vanish from the map. Dark Fleet is my graduation project on maritime awareness: what happens after a ship goes dark, and how to respond. More after the defence.

Dark Fleet · preview
known trackwhile dark

Research directions
exploring

Can a neural foundation model be personalised to one patient?

Personalised neural foundation models, with neurological and epilepsy-related applications in mind.

exploring

What should an explanation look like to the person who acts on it?

Explainable and human-centred AI, beyond feature-importance charts.

exploring

Can cameras watch an ecosystem continuously and honestly?

Computer vision for environmental monitoring, following on from RedSea-Lens.

exploring

How does a human stay in charge of a fast operational system?

AI for aviation safety, where alerts have to earn attention.

exploring

When should an agent retrieve, and when should it ask?

AI agents and RAG systems that know the limits of what they found.

exploring

How do text, images and time series reason together?

Multimodal intelligence and spatial-temporal prediction.

Let's build the next one.