Understand the need
I start with the people and the process: what it costs today, and what would change if it worked. Requirements, KPIs and business cases come before code.
Clinetta requirements, solar market studyThe same three steps, whether the project is an agent, a pipeline or a product.
I start with the people and the process: what it costs today, and what would change if it worked. Requirements, KPIs and business cases come before code.
Clinetta requirements, solar market studyAgents, pipelines and interfaces, with a clear line between what AI judges and what fixed rules calculate. Every claim should trace back to a source.
Ops Agent, CH Lead GeneratorDeployment, monitoring, guardrails and real measurement, so the system keeps working after it ships and the results can be checked.
Atlas, B2B Lead Scoring Agent
A harness I developed that turns an AI coding runtime into an autonomous operational agent, run entirely from plain-language files.
Claude Code · MCP · Markdown skills · CLI wrappers · Lifecycle hooks
Read case study
Pick an area on a map and get a ranked list of real buildings, each with a draft proposal.
Claude Haiku · Claude Sonnet · Google Places · OpenStreetMap · Microsoft footprints
Read case study
One shared clinic system that replaces a paper agenda and Word patient files.
React · Vite · Tailwind · FullCalendar · Fastify
Read case study
A competitor analysis agent and AI strategic analysis platform. Type a competitor's name and get a sourced battlecard, threat alerts and ranked strategic gaps.
Claude · MCP · React · Firebase · Firestore
Read case studyTurns a company name in Notion into an enriched, scored partner candidate with a ready-to-read brief.
Claude · Claude Cowork · Python · Notion API · Pappers
Read case study
A course assistant that answers from slides and diagrams, shows the figure, and cites the exact page.
LangGraph · LangChain · Gemini 2.5 Flash · Pinecone · FastAPI
Read case study
An explainable model that helps doctors estimate whether a patient's blood clot disease will come back.
Python · XGBoost · scikit-learn · SMOTENC · SHAP
Read case study16 projects across work and university. Filter by what interests you.
A harness I developed that turns an AI coding runtime into an autonomous operational agent, run entirely from plain-language files.
70–80% lower instruction-loading token cost
Pick an area on a map and get a ranked list of real buildings, each with a draft proposal.
10 automated steps per site
One shared clinic system that replaces a paper agenda and Word patient files.
1,064 automated tests
A competitor analysis agent and AI strategic analysis platform. Type a competitor's name and get a sourced battlecard, threat alerts and ranked strategic gaps.
12 research steps per competitorTurns a company name in Notion into an enriched, scored partner candidate with a ready-to-read brief.
216 prospects enriched and scored in production
A course assistant that answers from slides and diagrams, shows the figure, and cites the exact page.
2 modalities retrieved: text and images
An explainable model that helps doctors estimate whether a patient's blood clot disease will come back.
0.93 weighted F1 on the test set
Turns raw solar monitoring data into reports and justified product recommendations for each client.
80+ solar installations in the pipeline
Keeps negotiated prices current across suppliers and accounts, and shows exactly what failed to update.
4 suppliers on 4 different platforms
A Claude Code plugin that lets developers build on Odoo without the AI inventing APIs.
13 durable concepts in the guide
The Power BI business case that won me the internship, turning 23,368 installations into a two-speed marketing strategy.
23,368 installations analysedTurns free-form accident reports into structured JSON, and benchmarks three LLMs on the task.
3 models compared
A distributed ML pipeline that hunts rare fraud in 284,807 transactions, where only 0.17% are fraudulent.
284,807 transactions processed
Makes a ResNet-18 about 4× smaller and 2× faster, with almost no loss of accuracy.
44.9 → 11.4 MB model size after quantization
A convolutional autoencoder trained on real camera sensor noise, then tested on a different sensor.
32.4 dB mean PSNR on the test images
A real-time webcam demonstrator that detects and tracks moving objects, built on the classic Horn–Schunck method.
Real time on a live webcam feedAlso explored: Grover's quantum search algorithm in Qiskit, with single-solution, multi-solution and subset-sum applications.
Final-year engineering internship at a French solar company
Agents and automation on n8n
Workshops and talks, including with the AI Club at the Lebanese University.
A hands-on session that takes students from AI agent theory to a working retrieval agent on Telegram.
Research and a live proposal for one AI app covering the whole insurance journey, presented at the Phoenicia Hotel in Beirut.
A proposed 40-minute talk teaching students to use AI tools the way engineers use any other system.
Claude / Anthropic · n8n · LangGraph · LangChain · Prompt engineering · Agent architecture · MCP · RAG · LLMs · NLP · Hugging Face
See it in Ops Agent →PyTorch · TensorFlow · Computer vision · Machine learning · Spark / PySpark · XGBoost · SHAP & LIME
See it in MVTE Prediction →Python · Power BI · Power Query · Data pipelines · REST / GraphQL APIs · Data scraping · Excel
See it in Solar Analytics →React · Node.js · Firebase / Firestore · FastAPI · Docker · C# · Java · Matlab · Git
See it in Clinetta →Product strategy · Requirements & MoSCoW · UX/UI design · Market research · Agile / Scrum · Roadmapping · Go-to-market · Notion
See it in Clinetta →Workflow automation · B2B lead generation · Competitive intelligence · Business analysis · KPI design · HubSpot · Odoo
See it in Market Study →