Computer Vision · Edge AI · RAG / LLM Systems · Model Evaluation
Build the model → measure the behavior → connect the architecture → ship the system.
I'm an AI Engineer focused on turning models into complete, inspectable systems. My work spans computer vision, edge deployment, RAG / LLM engineering, model evaluation, APIs, data pipelines, and intelligent applications.
class AdhamElsayed:
role = "AI Engineer"
location = "Egypt"
focus = [
"Computer Vision",
"Edge AI",
"RAG / LLM Systems",
"Model Evaluation",
"Intelligent Applications",
]
systems = {
"Smart Basket": "edge vision + synchronized retail state",
"MedFlow": "evidence-grounded RAG + safety validation",
"Code AI Proctor": "visual inference + auditable exam workflow",
}
engineering_loop = "Data → Model → Evaluate → Serve → Product → Iterate"Featured Systems · Core Skills · Tech Stack · Evidence · Contact
Real-time retail product recognition running on Raspberry Pi 5, connected to basket-state logic, Firebase synchronization, and a Flutter application.
| Repository | Smart-Basket |
| Stack | YOLO · OpenCV · ONNX Runtime · Raspberry Pi 5 · Firebase · Flutter |
| Capability | Edge product detection + stabilized basket state + mobile synchronization |
| Validation | mAP@0.5 = 96.89% · mAP@0.5:0.95 = 84.79% |
flowchart LR
CAM[Pi Camera] --> PRE[OpenCV preprocessing]
PRE --> YOLO[YOLO detector]
YOLO --> EDGE[ONNX Runtime on Raspberry Pi 5]
EDGE --> STATE[Basket state tracker]
STATE -->|state changed| FB[(Firebase)]
FB --> APP[Flutter application]
EDGE -. runtime telemetry .-> MON[Runtime monitor]
Architecture notes
- Camera frames are prepared before inference with OpenCV.
- ONNX Runtime moves inference onto Raspberry Pi 5 instead of depending on a cloud model endpoint.
- Basket-state confirmation / hold behavior reduces unstable product-state changes.
- Firebase updates are written when state changes rather than continuously rewriting identical state.
- Flutter consumes the synchronized basket state at the application layer.
A thyroid-focused clinical AI prototype where retrieval, evidence sufficiency, generation, citation resolution, claim validation, numeric safety, and final policy are separate system layers.
| Repository | Medflow |
| Stack | FastAPI · BGE · ChromaDB · BM25 · RRF · Groq / GPT-OSS |
| Capability | Evidence-grounded answers with traceable citations and explicit safety routing |
| Retrieval | Precision@4 = 53.12% · Hit@4 = 87.50% · MRR ≈ 0.7031 |
flowchart TB
U[User] --> UI[MedFlow web UI]
UI --> API[FastAPI backend]
API --> GUARD[Input guardrail]
GUARD --> RISK{Risk class}
RISK -->|refuse / redirect| SAFE[Safe redirect]
RISK -->|allowed / caution| RET[Retrieval layer]
subgraph Retrieval
DENSE[Dense retrieval: BGE + ChromaDB]
SPARSE[Sparse retrieval: BM25]
FUSION[Reciprocal Rank Fusion]
DENSE --> FUSION
SPARSE --> FUSION
end
RET --> DENSE
RET --> SPARSE
FUSION --> GATE{Evidence sufficiency gate}
GATE -->|block| ABSTAIN[Abstain]
GATE -->|pass / downgrade| PACK[Evidence packaging E1..E4]
PACK --> LLM[Grounded LLM]
LLM --> CITE[Citation engine]
LLM --> CLAIM[Claim validator]
LLM --> NUM[Numeric / dosage validator]
CITE --> POLICY{Final safety policy}
CLAIM --> POLICY
NUM --> POLICY
POLICY -->|safe| ANSWER[Answer]
POLICY -->|caution| CAUTION[Answer with caution]
POLICY -->|insufficient| ABSTAIN
POLICY -->|unsafe| SAFE
flowchart LR
PDF[Medical PDFs] --> EXT[Text extraction]
EXT --> CLEAN[Cleaning]
CLEAN --> CHUNK[Token-aware chunking]
CHUNK --> META[Metadata enrichment]
META --> EMB[BGE embeddings]
EMB --> DB[(ChromaDB)]
Architecture notes
- Dense and sparse retrieval remain separate before fusion.
- Evidence sufficiency can stop generation before the LLM is allowed to answer.
- Citation resolution is deterministic and separated from claim support.
- Numbers, units, dosages, thresholds, and durations receive stricter validation.
- The final policy can answer, caution, abstain, or redirect.
The reported values are engineering evaluation metrics, not clinical validation or a medical-accuracy claim.
A self-hosted exam platform connecting webcam inference to organization management, quizzes, incident persistence, teacher review, grading, and student appeals.
| Repository | code-ai-proctor |
| Stack | YOLO · PyTorch · FastAPI · SQLAlchemy · HTML/CSS/JS |
| Capability | Local visual inference connected to an auditable exam workflow |
| AI task | Binary classification: cheating vs normal |
flowchart LR
CAM[Browser webcam] --> FRAME[Canvas snapshot]
FRAME --> API[FastAPI /api/predict]
API --> CROP[Center crop]
CROP --> YOLO[YOLO classifier]
YOLO --> TEMP[Temporal logic]
TEMP -->|normal| CONT[Continue exam]
TEMP -->|confirmed alert| INCIDENT[Cheating incident]
INCIDENT --> DB[(SQLAlchemy database)]
DB --> TEACH[Teacher review]
TEACH --> GRADE[Grades / CSV export]
DB --> APPEAL[Student appeal]
flowchart TD
ORG[Organization] --> ACC[Teacher / student accounts]
ACC --> CLASS[Teacher creates class]
CLASS --> QUIZ[Teacher creates quiz]
QUIZ --> JOIN[Student joins class]
JOIN --> ATTEMPT[Quiz attempt]
ATTEMPT --> PROCTOR[Live AI proctoring]
PROCTOR --> REVIEW[Incident review]
REVIEW --> RESULT[Grade / appeal workflow]
Architecture notes
- YOLO inference runs locally with CUDA when available and CPU fallback otherwise.
- Webcam frames are center-cropped before classification to avoid aspect-ratio distortion.
- Batch probability and streak logic reduce single-frame instability.
- Confirmed incidents are attached to the active quiz attempt and persisted for review.
- Teacher notifications, acknowledgement, grading, CSV export, and student appeals extend the system beyond inference.
Big-data style churn analysis and binary classification using Spark DataFrames and MLlib.
| Repository | Telco-Customer-Churn-Project |
| Stack | PySpark · Spark DataFrames · MLlib · Google Colab |
| Task | Predict whether a telecom customer will churn |
| Model | Logistic Regression |
flowchart LR
CSV[Telco churn CSV] --> SPARK[Spark DataFrame]
SPARK --> EDA[Analysis]
EDA --> CAT[Categorical columns]
CAT --> IDX[StringIndexer]
IDX --> OHE[OneHotEncoder]
OHE --> FEAT[Feature vector]
FEAT --> LR[Logistic Regression]
LR --> EVAL[Accuracy · F1 · Precision · Recall · AUC]
EVAL --> CM[Confusion matrix]
A compact exploratory analysis connecting study behavior, attendance, previous performance, final score, and pass/fail status.
| Repository | Student-Performance-Analysis |
| Stack | Python · Pandas · Matplotlib · Google Colab |
| Focus | Exploratory student-performance analysis and visualization |
flowchart LR
DATA[Student data] --> DF[Pandas DataFrame]
DF --> QUALITY[Null checks]
QUALITY --> STATUS[Derived Pass / Fail]
STATUS --> STATS[Descriptive statistics]
STATS --> REL[Hours Studied vs Final Score]
REL --> VIZ[Matplotlib visuals]
STATUS --> DIST[Pass / Fail distribution]
DIST --> VIZ
Inspect Student Performance Analysis →
01 Understand the failure mode before adding complexity.
02 Measure behavior before calling a component better.
03 Tie every metric to its exact evaluation context.
04 Separate retrieval, generation, validation, and product workflow.
05 Prefer inspectable architecture over black-box demos.
06 Move notebook → inference → API → product.
| Project | Measured / inspectable evidence |
|---|---|
| Smart Basket | mAP@0.5 = 96.89% · mAP@0.5:0.95 = 84.79% |
| MedFlow | Precision@4 = 53.12% · Hit@4 = 87.50% · MRR ≈ 0.7031 |
| Code AI Proctor | Local inference · temporal stabilization · incident persistence · review workflow |
| Telco Churn | Accuracy · F1 · Precision · Recall · AUC · Confusion Matrix |
| Student Performance | Descriptive statistics · relationship analysis · pass/fail visualization |
| Education | B.Sc. Artificial Intelligence Engineering — Mansoura University · Sep 2023 – Expected Feb 2027 |
| Academic standing | GPA 3.13 / 4.00 · Grade B+ |
| Training | Digital Egypt Pioneers Initiative — Microsoft AI & Data Science / Machine Learning Engineering · Sep 2025 – Jul 2026 |
| Location | Egypt 🇪🇬 |
📊 GitHub activity
- Computer Vision & Edge AI — detection / classification, efficient local inference, model-to-device integration.
- RAG / LLM Systems — retrieval quality, evidence sufficiency, grounding, citation and claim validation.
- Evaluation — measurable behavior, failure modes, and system-level iteration.
- AI Product Engineering — model → API → workflow → usable product.
Adham Elsayed · AI Engineer · Egypt 🇪🇬
Build the model · Measure the behavior · Ship the system