AI Engineer · Applied AI Systems

Marwan Al‑Masrat

Physics-trained AI Engineer building measurable, production-minded systems across LLM fine-tuning, RAG, knowledge graphs, NLP, and computer vision - from model evaluation to APIs and observability.

LLM Fine-Tuning RAG & Retrieval Knowledge Graphs Production AI Computer Vision
Scroll to explore
Professional Profile

AI engineering,
grounded in first principles

Physics-trained AI Engineer focused on end-to-end applied AI: LLM fine-tuning, retrieval-augmented generation, knowledge graphs, NLP, computer vision, and predictive machine learning.

I turn experiments into measurable systems using rigorous evaluation, vector databases, FastAPI, Docker, and observability - with clear evidence of quality, latency, and reliability.

Alongside technical delivery, I founded and led Sufra Alqalam, a 110-member initiative awarded 5,000 JOD by the Jordanian Ministry of Culture.

22
Selected AI Projects
95%
RAG QA Accuracy
110
Members Led
24W
Applied AI Program
Project Portfolio

Applied AI systems,
from research to production

22 projects across LLM fine-tuning, retrieval, knowledge graphs, NLP, computer vision, predictive ML, and AI infrastructure - organized by specialization for fast review.

Fine-Tuning & Agentic Systems02 projects

Flagship work first: a reproducible domain fine-tuning lifecycle and a collaborative agentic platform with controlled actions.

RAG & Retrieval Engineering05 projects

Retrieval systems ordered from full multimodal RAG through focused experiments in hybrid search, routing, re-ranking, and document QA.

Multimodal RAG for Scientific Papers
Combined ColPali visual page retrieval with semantically chunked text, Weaviate hybrid search, and cross-encoder re-ranking. On a 20-question paper benchmark, answer accuracy improved from 80% to 95%.
WeaviateColPaliHybrid SearchStreamlit
Explore repositorygithub.com
Hybrid Retrieval Benchmark
Benchmarked BM25, MiniLM dense, and Weaviate hybrid retrieval across factoid and paraphrastic queries using Recall@5, Recall@10, and MRR. Dense reached 0.900 Recall@5; hybrid reached 0.983 Recall@10.
WeaviateBM25MiniLMRecall / MRR
Explore repositorygithub.com
Adaptive Query Router
Classifies factoid, paraphrastic, and mixed queries, then dispatches them to BM25, dense, or hybrid retrieval. The evaluated rules reached 85% routing accuracy and 0.93 Recall@5 on factoid queries.
Query RoutingBM25Dense RetrievalHybrid Search
Explore repositorygithub.com
Cross-Encoder Re-Ranking Benchmark
Added an MS MARCO MiniLM cross-encoder to a hybrid Weaviate pipeline and measured both retrieval quality and CPU latency. The study exposed a 194x overhead and documented when re-ranking should be reduced or skipped.
Cross-EncoderRe-RankingLatency ProfilingWeaviate
Explore repositorygithub.com
AI RAG Assistant with LangChain
Built a document-question-answering pipeline with PDF parsing, recursive chunking, IBM watsonx embeddings and generation, ChromaDB retrieval, and a Gradio upload-and-chat interface.
LangChainIBM watsonxChromaDBGradio
Explore repositorygithub.com
Knowledge Graphs & Semantic AI04 projects

Schema-aware systems for translating language into graph queries, validating claims, combining graph structure with vectors, and querying RDF knowledge.

Natural Language to Cypher Engine
A deterministic, schema-bounded mapper for 15 question shapes with intent detection, slot extraction, parameterized Cypher, Neo4j execution, and explicit rejection of unsupported requests.
Neo4jCypherIntent MappingQuery Safety
Explore repositorygithub.com
Knowledge Graph Claim Critic
A three-stage critic that labels claims as supported, entailed, contradicted, or unsupported using direct checks, hierarchy traversal, and schema constraints. It achieved perfect precision and recall on a 40-claim evaluation set.
Neo4jKG ReasoningEntailmentSchema Validation
Explore repositorygithub.com
GraphRAG Hybrid Retrieval
Combines a Neo4j vector index with one-hop graph expansion and structural score fusion. The pipeline reached 0.812 macro Recall@10 across the evaluation set while exposing where categorical graph signals improve vector search.
GraphRAGNeo4j Vector IndexMiniLMScore Fusion
Explore repositorygithub.com
Natural Language to SPARQL Dispatcher
Maps supported natural-language publication intents to auditable SPARQL queries and executes them against an RDF ontology hosted in Apache Jena Fuseki.
SPARQLRDFApache JenaDocker
Explore repositorygithub.com
AI Services, Orchestration & Reliability04 projects

Production-oriented AI infrastructure covering service routing, failure isolation, distributed tracing, metrics, dashboards, and load profiling.

Multi-Service AI Coordinator
A Docker Compose system that classifies requests, fans out concurrently to NLP, knowledge-graph, and RAG services, aggregates responses, and degrades gracefully during partial failures. Its coordinator suite passes all six tests.
FastAPIMicroservicesAsyncIODocker Compose
Explore repositorygithub.com
Distributed AI Router & Observability
A three-service FastAPI router with per-service Prometheus metrics, structured logs, and propagated request IDs. It classified all 15 held-out questions correctly and passed all 16 automated tests.
FastAPIPrometheusRequest TracingService Routing
Explore repositorygithub.com
FastAPI Monitoring & Observability Stack
Instrumented an NLP, knowledge-graph, and RAG API with request IDs, structured JSON logs, Prometheus counters, latency histograms, in-flight gauges, a Grafana dashboard, grounding checks, and tests.
FastAPIPrometheusGrafanaDocker
Explore repositorygithub.com
AI Service Load Profiler
An HTTPX load generator that ramps concurrency, captures p50/p95/p99 and error rate, identifies the latency knee, and produces a chart-backed report. The measured knee appeared at load 5 after a 62.6% p95 increase.
HTTPXLoad Testingp95 / p99Matplotlib
Explore repositorygithub.com
NLP Evaluation & Generative AI02 projects

Evaluation-first NLP work and an applied reasoning pipeline that converts large-scale customer feedback into structured recommendations.

NLP Evaluation Lab
Evaluates extractive QA and abstractive summarization with EM, Token-F1, ROUGE, normalization traces, batch comparisons, numeric-faithfulness checks, and an auditable fine-tune decision matrix.
TransformersEM / F1ROUGEStreamlit
Explore repositorygithub.com
GenAI Reasoning for Customer Reviews
Analyzed more than 9,400 Yelp reviews through sentiment processing and a GenAI reasoning layer, then converted the findings into structured business insights with Arabic conversational analysis.
NLPOpenRouterPrompt EngineeringData Analysis
Explore repositorygithub.com
Computer Vision & Deep Learning02 projects

Applied deep-learning work for domain-specific object detection and renewable-energy forecasting, each backed by measured outcomes.

Jordanian Football Object Detection
Prepared and annotated roughly 300 match frames, fine-tuned YOLOv8 Medium and Large variants across multiple experiments, and ran full-video inference for players, balls, referees, and goalkeepers.
YOLOv8RoboflowUltralyticsVideo Inference
Explore repositorygithub.com
Solar Power Generation Predictor
Engineered weather features and trained a Keras 128-64-1 regression network with early stopping, reaching R-squared 0.79 and RMSE 439.7 kW on held-out data, then deployed it with Streamlit.
TensorFlowKerasFeature EngineeringStreamlit
Explore repositorygithub.com
Predictive ML & Model Selection03 projects

Classical machine-learning projects centered on honest evaluation, tuning, reproducible comparison, and interpretable performance diagnostics.

Telecom Churn Model Comparison CLI
A production-style command-line tool that compares six model configurations with stratified cross-validation and saves metrics, diagnostic plots, and the best serialized model.
Scikit-learnCross-ValidationCLIJoblib
Explore repositorygithub.com
Hyperparameter Tuning & Nested CV
A Random Forest and Decision Tree study using GridSearchCV and nested cross-validation to separate tuning from honest evaluation. The selected forest scored 0.4899 inner-CV F1 versus 0.4806 on outer folds.
GridSearchCVNested CVRandom ForestModel Selection
Explore repositorygithub.com
Mobile User Behavior Classifier
Built a K-nearest-neighbors classifier for five mobile-usage behavior classes from 700 records, reaching 90.7% validation accuracy and a 90.7% F1-score.
KNNClassificationPandasModel Evaluation
Explore repositorygithub.com
Contact & Availability

Open to AI engineering
opportunities

Seeking AI Engineer roles where I can build and evaluate reliable LLM, retrieval, NLP, computer vision, and production ML systems. The fastest ways to reach me are below.

Available
for Work→ AI / ML