My Stack
TOOLS I BUILD WITH
Manual crop health monitoring is time-consuming and prone to human error, leading to avoidable yield loss.
Deployed fine-tuned CNN models onto Raspberry Pi edge hardware for autonomous real-time field diagnostics.
Free-text clinical notes in emergency care often miss critical guideline parameters, while LLM auditors introduce latency, high API costs, and dangerous hallucination risks.
Built a lightweight, rule-grounded Python audit engine with smart negation parsing, 12 unit tests, and an interactive Streamlit UI.
Multi-tenant support agent builder that resolves customer queries via a router→retrieve→grade→rewrite→answer pipeline with human-in-the-loop escalation — cuts manual ticket triage by 70% while guaranteeing zero-hallucination answers.
Telegram-based AI meeting and relationship intelligence agent that builds persistent context around people and organizations — featuring scheduled pre-meeting briefings, post-meeting commitment extraction, and a deterministic approval-gated MCP architecture where the LLM is never the single source of truth.
Real-time video intelligence and multi-object tracking platform transforming unstructured surveillance feeds into queryable spatial intelligence — featuring YOLOv8 + ByteTrack (300+ FPS), 3×3 metric planar homography, deterministic spatio-temporal event rules, and a tool-using Vision LLM agent.
Full-stack AI nutrition platform estimating calories and macronutrients directly from food photographs — built on Next.js 15 and FastAPI with a production Clean Architecture (Controllers → Services → Providers → Repositories), asynchronous batch image analysis via asyncio.gather, and Google Gemini Flash multimodal vision.
Multilingual RAG system built on FastAPI and pgvector with a Streamlit UI — featuring swappable LLM providers (Cohere, OpenAI, Ollama) and grounded, source-attributed answers in Arabic and English.
From-scratch PyTorch implementation of the Transformer architecture converging to 1.61 training loss, featuring self-attention, sinusoidal positional encoding, and causal masking.
Modular PyTorch implementation of PointNet achieving 86% overall test accuracy on ModelNet, featuring dual T-Nets, shared MLPs, and orthogonal regularization.
Open to collaborations, freelance projects, and full-time AI engineering opportunities.
Fill out the form below and I'll get back to you within 24 hours.
