Applied AI

Model-driven software embedded in a defined workflow with review, safety, and operational constraints.

Projects

• Multilingual Medical Prescription OCRA CLI and FastAPI service that uses Gemini to extract medicine names from Arabic, English, and French prescription images, with an optional medicine database.
• AI Customer Support CouncilA self-hosted internal admin console for synthetic B2B support intake and triage, built as a technical-assessment deliverable for Valsoft Corporation. It is not a customer-facing platform.
• Hantawatch — outbreak situational-awareness dashboardA deployed situational-awareness dashboard for the MV Hondius hantavirus outbreak of April–May 2026, built from public sources only: a country choropleth, status-colored case events, a linked event feed, indicators, and a news ticker.
• Basira — retinal screening decision supportA prototype web platform for eye clinics in Lebanon in which three independent retinal-image models each analyze a fundus or OCT image, a consensus engine reports agreement, urgency, and a referral suggestion, and the reviewing doctor confirms or overrides before a report is issued.
• Imagen — learning a photographer’s edit from RAW bracketsA freelance pipeline for a French real-estate photography agency that ingests bracketed RAW sessions, decodes them through a color-accurate DNG pipeline into 16-bit linear frames, pairs them with the photographer’s edited JPEGs, and trains an HDRNet model to replicate the editing style.
• Lumiscan dashboard — tracking skin lesions across scansA Next.js and PostgreSQL platform that lets a clinic record patients, lesions, and successive scans from a lesion-scanning device, view metric trends over time, and generate narrative summaries, with strict organization scoping and a stubbed device-ingestion API.

Writing

• Extracting medicine names from multilingual prescriptionsA small document-intelligence service for Arabic, English, and French prescriptions—and an example of why medical OCR needs explicit human verification.
• Building a local-first AI support triage councilAn internal support workflow where an LLM proposes structure, deterministic rules enforce policy, and a human keeps the final say.
• Building an outbreak dashboard from public sourcesAggregating what agencies and newsrooms already publish into one map is useful only if every claim stays linked to where it came from.
• Designing a council of models for retinal screeningThree foundation models that disagree in the open are more useful to a clinic than one model that is confidently wrong.
• Why color science comes before the modelTo teach a network a photographer’s edit, the training pairs must be linear, calibrated, and exact. Most of the work is in the decode.
• Making tenant isolation the only pathFor a clinical platform, the dominant risk is one organization seeing another’s patients. The architecture should make that mistake hard to write.